138results about How to "Solve missing detection" patented technology

The invention discloses a quality safety tracing method for food processing enterprises. The quality safety tracing method includes the steps that a laser coder generates random tracing codes according to a coding rule when a laser marking machine in a production line device is dynamically read in an on-line tagging mode; the laser marking machine carries out tracing mode marking on products on a production line, and two-dimension codes and plain codes are all marked on packages; a fixed reader is used for scanning and reading designated tagged code portions of the products; the two-dimension codes are successfully scanned and read, the tracing codes are recorded, and relevance between the tracing codes and commodity attribute information is completed; if the two-dimension codes fail to be scanned, the two-dimension codes will be removed through a removing device; the commodity tracing code information and the commodity attribute information are recorded into a relational database; after the products are packaged, package code data and product tracing code data are related; the products are stacked, stacking codes and tagged stacking codes are generated, and the stacking codes and the product package code data are related; the stacking codes are scanned, storage of the products is completed, delivery orders and the stacking codes are scanned to complete delivery, and the database records delivery information.

The invention relates to a PDA (personal digital assistant)-based realization method for intelligent patrol of electric equipment. The method is characterized by comprising the following steps: downloading patrol task information to a handheld PDA from an office PC (personal computer) by a patrol staff, patrolling predefined equipment patrol items one by one through an RFID (radio-frequency identification) electronic label by the handheld PDA during outdoor patrolling, and recording data content; after the patrol task is finished, submitting a patrol result to a server through a patrol management host. According to the method, the patrol quality of the electric field equipment can be improved, the problem of non in-place patrolling, missing patrolling, non on-time patrolling or the like is solved, the patrolling efficiency is improved, and the situations of low efficiency, high possibility of missing items and errors during manual reporting in the past are avoided.

The invention discloses a multilayer quotation recommendation method based on a literature content mapping knowledge domain, and belongs to the field of information recommendation and intelligent information processing. The method comprises the following steps: firstly, obtaining the query requirement of a user, wherein the query requirement consists of the key words of the title and the digest of a thesis which needs to recommend a quotation thesis or quotation literature; then, on the basis of the literature content mapping knowledge domain, expanding and querying a retrieval word, wherein the mapping knowledge domain consists of the research object word and the research behavior word node of the literature, and edges which express various semantic relations including synonymy, synonym, an up and down position, part-whole, juxtaposition and the like; and finally, constructing the inverted index of the literature in a data set, selecting a candidate quotation, calculating the similarity between the candidate quotation and query, and adopting a gradient progressive regression tree to carry out quotation recommendation. The method carries out multilayer quotation recommendation on the basis of the literature content mapping knowledge domain, enlarges the range of the candidate quotation, accurately expresses the research object and contents of the thesis, improves efficiency for users to obtain a relevant literature and has a wide application prospect.

The invention discloses a printing image defect detection method, which comprises the following steps: carrying out real-time image model learning in a gray scale region and a gradient region by aiming at specific images on a large image; comparing the large image to be detected to the a gray scale region model and a gradient region model established during learning; realizing small-dimension strong-contrast defect detection through Blob cluster analysis; when no small-dimension strong-contrast defect is detected, dividing the large image into sub regions, and respectively calculating image integrated features; adopting a variable threshold method for carrying out threshold division on each sub region; carrying out Blob cluster analysis on divided images; and realizing the large-area weak-contrast defect detection. Compared with the prior art, the invention does not rely on a reference template, is not sensitive on the real-time imaging brightness change, can overcome the defects of missing detection, error detection and poor self adaptation of the template by using a reference template comparison detection method, and can simultaneously solve the problems of eliminating tiny wrinkles and blackspots in printing and papermaking industries.

The invention discloses a bamboo strip defect online detection method based on machine vision. A camera is used for photographing the four surfaces of a bamboo strip continuously along the length direction of the bamboo strip to obtain the gray images of the surfaces of the bamboo strip. The image treatment is implemented on every captured frame of digital images and whether defects exist on the small section of the bamboo strip in the frame of the images is judged according to the bamboo strip defect detection method. If the small section of the bamboo strip in the frame of the images is provided with detects, the whole bamboo strip is considered to be provided with detects and simultaneously an instruction of the existing defects is sent to a sorting mechanism. If small sections of the bamboo strip in all frames of the images of the whole bamboo strip are provided with no detects, the bamboo strip is considered to be provided no detects. Simultaneously, an instruction with no detects is sent to the sorting mechanism. The method can implement the online detection on the defects of bamboos with different color shades and simultaneously detect the four surfaces along the length direction of the bamboo strip.

The invention discloses a circumferentially-distributed crawler wheel type pipeline detection robot capable of actively adapting to pipe diameter changes. The robot comprises a crawler wheel travelingmechanism and a detection mechanism connected to the front end of the crawler wheel traveling mechanism through a turning mechanism, the crawler wheel traveling mechanism comprises a cylindrical mainbody sleeve, multiple crawler wheels controlled by a crawler driving motor are evenly distributed on the periphery of the cylindrical main body sleeve, the crawler wheels are installed on a ball screw through a pressing adjusting crank sliding block mechanism, and the ball screw is fixed to the outer wall of the cylindrical main body sleeve. The detection mechanism comprises a cylindrical detection barrel, an ultrasonic probe and a wide-angle camera stretch out of a detection barrel front end cover of the cylindrical detection barrel, the front end of the ultrasonic probe is provided with anultrasonic reflector, an electrical control system is arranged inside the cylindrical main body sleeve, and the electrical control system is connected with a battery. The circumferentially-distributedcrawler wheel type pipeline detection robot is good in movement controllability, outstanding in climbing and obstacle crossing ability and capable of actively adapting to pipe diameter working conditions.

The invention discloses a method and a device for automatically detecting the boundary or region of beam-limiting device. The method mainly comprises a reading step for reading an image acquired through an X-ray imaging system, a detection step for detecting transition points between a foreground and a background as well as between the foreground and an organization on the X-ray image, and a transformation step for detecting boundary straight lines between the foreground and the background as well as between the foreground and the organization by utilizing the detected transition points through HT or RT transformation. The detecting method and the detecting device have the advantages of simple calculation, high detection ratio and capability of effectively solving detection omission and false positive problems.

The invention discloses a method and a system for locating a moving target in a mine based on binocular CCD vision. The system for implementing the method comprises surface equipment and an underground device. The surface equipment comprises a base station controller, a location server, an Ethernet switch and a monitoring terminal. The underground device comprises a mine intrinsically safe base station and a wireless radio frequency identification tag. The method for implementing the system comprises the following steps: (1) using an RSSI algorithm to estimate the location of a moving target; (2) carrying out matching feature extracting and training on the wireless radio frequency identification tag of the moving target; (3) carrying out feature matching on images after sampling through an ORB algorithm; (4) using a binocular CCD vision sensor to stereoscopically calibrate the left and right images of the moving target; (5) getting an internal and external parameter matrix of the binocular CCD vision sensor, and calibrating the world coordinates of the moving target; and (6) carrying out location correction on the world coordinate information of the moving target, and getting the accurate location information of the moving target in a mine. The problem on how to accurately locate a moving target in a mine NLOS environment is solved. The reliability and robustness of the system are improved.

The invention is suitable for the technical field of automatic control, and provides a production line device used for similar multistage assembly product code assigning and tracing. A first bar code reader 5 is used for reading EAN-13 codes of single stage products after separation, a first rejection device 6 is used for rejecting the single stage products with the reading EAN-13 codes, not involved in current work order information, a first obstruction cylinder 7 is used for obstructing the single stage products under a laser coding machine 8, the laser coding machine 8 is used for printing two-dimensional codes on the upward sides of the single stage products, a first pushing cylinder 9 is used for pushing the obstructed single stage products periodically, an interval separation device 10 is used for separating the single stage products moving again at equal intervals according to a setup parameter, a two-dimensional code collector 11 is used for reading the two-dimensional codes of the single stage products, and a second rejection device 12 is used for rejecting the single stage products with the two-dimensional codes which cannot be read successfully. Through operations of an IPC and a PLC device, the parameter setting is finished, so that the production line device can produce special-type or special-stage finished products as required, and has the universality, and the cost is reduced.

The invention relates to a vehicle detection method based on monocular vision and laser radar fusion. The method comprises the following steps of S1, obtaining an image feature map; S2, obtaining a point cloud feature map; S3, respectively extracting a point cloud feature vector flidar and an image feature vector fRGB from the point cloud feature map and the image feature map; S4, performing feature fusion on the point cloud feature vector flidar and the image feature vector fRGB to obtain a fusion feature fL; S5, obtaining a 3D bounding box of the vehicle according to the fusion feature fL, and obtaining corresponding category parameters; S6, removing the overlapped 3D bounding boxes to obtain the final 3D bounding boxes to complete vehicle detection. Compared with the prior art, the method has the advantages that the problems that the position of the vehicle is difficult to effectively estimate through monocular vision and missing detection is possibly caused by laser radar due to long-distance point cloud sparseness are solved, and the three-dimensional target detection effect on the vehicle is further improved.

The present invention discloses a pedestrian detection method based on semantic segmentation information, and relates to the field of a pedestrian detection method based on a neural network. The pedestrian detection method based on semantic segmentation information comprises the steps of: 1: inputting original RGB images in training set samples into a backbone network, inputting corresponding semantic segmentation images into branch networks, and setting a loss function of the whole network to complete training; 2: inputting the original RGB images in THE training set samples into the backbonenetwork which completes training to perform convolutional feature extraction and generate a multi-layer feature map; 3: inputting the multi-layer feature map into an area generation network which completes training to perform pedestrian candidate frame extraction and generate a pedestrian candidate area; and 4: performing classification and location of the pedestrian candidate area by a classification regression network which completes training to output detection result images including a pedestrian position surrounding frame. The problem is solved that pedestrians and backgrounds are difficult to be distinguished in a low-resolution condition to cause low detection precision in the current pedestrian detection, and the precision of pedestrian detection is improved in the low-resolutioncondition.

The invention discloses a multi-angle human face detecting method based on weighting of deformable components. The method includes a model building process and a human face detecting process. Different weights are set on different components of a human face according to different contributions of the different components to a detection effect, and effects on important components can be sufficiently utilized assuredly. The problem in terms of large-angle human face detection is resolved by a multi-model fusion method, omission ratio is decreased by means of reducing a threshold value in a detection process, and false detecting ratio is decreased by means of a skin color confirming mode in a model result fusion process. The multi-angle human face detecting method effectively resolves problems of omission and false detection in the human face detecting process based on the components, and can be widely applied to the fields of safety detection, identity authentication, intelligent transportation and the like.

The invention relates to a box camera and ball head camera linkage-based road side parking detection system. A minimal detection unit of the system consists of a box camera and a ball head camera of which both have intelligent video analysis capability; the minimal detection unit is formed either via integration or combination of the box camera and the ball head camera; combination of the box camera and the ball head camera requires installation of an exchanger device which enables the two to access a network at the same time; the whole system consists of N minimal detection units; rods that are L meters in height are erected every X meters on a sidewalk of a side face of a roadside parking area; one minimal detection unit is installed on each rod. According to the system and the method, composite application of a plurality of intelligent video detection technologies such as vehicle movement detecting and tracking technologies, vehicle license plate recognition technologies, personneldetection technologies and the like is adopted; road side parking scenes, vehicle license plate information, vehicle characteristic information, conductor and driver information and parking timing information are bound together; timing and charging accuracy and reliability and charging basis validity can be improved; problems of non-transparent timing and charging, payment refusal and the like ina conventional solution can be effective solved.

The invention discloses a FAST-AIC-algorithm-based automatic first-arrival collecting method for micro-seismic signals. The method includes: providing a new identification method of micro-seismic events according to the fact that the propagation path and the travel time information of hydraulic fracturing perforation signals are similar to those of the micro-seismic signals; using a hydraulic fracturing wave detector arrangement manner and a single-channel signal local extreme value to design a precise time window near a micro-seismic triggering moment; using Curvelet conversion to filter data in the precise time window; performing first-arrival collecting in the precise time window through a FAST-AIC algorithm, wherein the FAST-AIC algorithm is obtained by optimizing and converting a traditional AIC algorithm, the linear combination manner of the arithmetic sum and the quadratic sum of a discrete real number sequence is obtained, repeated calculation is reduced, and calculation efficiency is increased greatly. The new method is used to perform first-arrival collecting on analog data and actual hydraulic fracturing micro-seismic monitoring data to verify the feasibility and correctness of the new method. Compared with the first-arrival collecting result of a conventional micro-seismic method, the new method has evident advantages.

The invention discloses a multiframe accumulated noncoherent MIMO radar detection and positioning method, and belongs to the technical field of radar signal processing. The method directs at a weak target in a low signal to noise ratio environment, threshold processing is not performed on single frames, a log-likelihood ratio function is used as a value function, multiframe accumulation is performed on value functions of a few continuous frames according to a state transition relation of the target, information of target multiframe is fully utilized, and after joint processing of multiframe data, a detection result and an estimation position of the target are announced at the same time. The multiframe accumulated noncoherent MIMO radar detection and positioning method solves the problems of large positioning error and missing detection of the weak target in the low signal to noise ratio environment which are likely to occur, and compared with an existing maximum likelihood estimation method, can obtain better detection probability and higher positioning accuracy.

The invention relates to an integral ultrasonic guide wave detecting method for a barrel of a pressure vessel. The integral ultrasonic guide wave detecting method includes steps of 1, calculating a phase speed and group speed frequency dispersion chart of longitudinal or peripheral guided waves; 2, determining a range of detecting guide wave excitation frequency; 3, determining the best detecting guide wave frequency and a guide wave incident angle; 4, utilizing a common ultrasonic flaw detector and setting sound speed options of the ultrasonic flaw detector; and 5, correctly connecting a probe with the detector, placing the probe onto the barrel of the vessel to axially or peripherally scan, and utilizing reflecting echo of a weld joint of the barrel of the vessel to adjust sensitivity or processing an artificial reflector to determine the detection sensitivity according to defects required to be detected until detection is finished. The method can be used for efficiently and comprehensively detect the defects of the barrel of the pressure vessel.

The invention discloses an oligonucleotide, a kit and a method for detecting novel coronavirus SARS-CoV-2, and relates to a primer and a probe for designing specificity aiming at Orf1ab, N and S generegions in a novel coronavirus genome. After nucleic acid in a sample is rapidly extracted, a one-step fluorescent RT-PCR technology is adopted to detect novel coronavirus nucleic acid. The method issimple to operate, and three target genes of the novel coronavirus are detected at the same time, so that missing detection can be avoided.

The invention discloses an intelligent test system and method for a metering device circuit board, belongs to the circuit board detection technology field and solves a problem of low efficiency caused by manual detection existing in a metering device circuit board in the prior art. The intelligent test system comprises a programming device used for realizing programming of function programs into a to-be-tested circuit board, a display performance test device used for testing display performance of the to-be-tested circuit board, a function test device used for testing functions of the to-be-tested circuit board, an identification device used for identifying the to-be-tested circuit board according to test results of the display performance test device and the function test device, a circuit board conveying mechanism which is connected with the programming device and is used for conveying the to-be-tested circuit board to the programming device, and a main control unit used for controlling operation of the programming device, the display performance test device, the function test device, the identification device and the circuit board conveying mechanism. The intelligent test system is advantaged in that the intelligent test system is applicable to circuit board tests of various types of intelligent gas meters.

Embodiments of the invention disclose a method for realizing single-source and multi-view security inspection, the method is used in the field of the security inspection. The method comprises that: a ray source emits a radiation ray beam to carry out a perspective of the current angle or the current position for an object requiring inspection; a detection device detects the ray generated due to the perspective of the object requiring inspection, and adopts the ray as projection data; the projection data is treated to reconstitute the image of the object requiring inspection, and to judge whether the image achieves a security inspection purpose; if the image does not achieve the security inspection purpose, a rotation angle is arranged, the ray source and the detection device rotate to the rotation angle or the above steps are repeated performed until the purpose of the security inspection is achieved. The embodiments of the invention further disclose a system for realizing the single-source and multi-view security inspection. With the system, the multi-view, omnibearing and positioning scan within 360 degrees can be completed by the rotation, the image of the detected object can be rapidly reconstituted. The method and the system are applicable for any occasions requiring safety inspections.

The invention relates to the technical field of computer vision object detection, more specifically, to a deep network pedestrian detection method based on shallow feature fusion guidance. On the basis of CNN detection algorithm, a pedestrian detection method in deep network based on shallow feature fusion is proposed. Through the fusion of directional gradient histogram, improved texture featuresand depth network features, accurate pedestrian features are obtained, which are guided by shallow features with depth learning as the core, and complementary advantages of shallow learning and depthlearning are realized. The results show that the improved method has good performance in detection accuracy and speed. the method effectively solves the technical problem of missing detection in complex scenarios and small targets.

The invention relates to a testing method of a machine vision based micro-sound film concentricity online testing system. The machine vision based micro-sound film concentricity online testing system comprises an optical imaging subsystem, a digital image processing subsystem, a motion control subsystem, a sound film eliminating subsystem, an image capture card and a motion control card, wherein signals of the optical imaging subsystem are sent to the image capture card through an interface, and the image capture card captures a frame of sound film gray scale image data and sends the data to the digital image processing subsystem to be processed and judged; a signal output end of the digital image processing subsystem is connected with a signal input end of the motion control card, and a signal port of the motion control card is respectively connected with the motion control subsystem and the sound film eliminating subsystem; and the motion control subsystem and the sound film eliminating subsystem are intensively controlled by the motion control card. The machine vision based micro-sound film concentricity online testing system realizes the precise positioning of sound films and the elimination of unqualified sound films by the motion control card controlling a stepper motor, a servo motor, a sensor, an eliminating device and the like, thereby realizing the high-precision online real-time testing of the sound films.

The invention discloses a method and device for detecting a target object, and relates to the technical field of computers. A specific embodiment of the method comprises the following steps: extracting a sample picture set from a historical video file recorded in a to-be-detected area; performing target object detection on the sample picture set based on a first confidence coefficient threshold value and a second confidence coefficient threshold value which are different in size to obtain sample data sets under different confidence coefficient thresholds; determining an invalid region by comparing the sample data set under the first confidence threshold value with the sample data set under the second confidence threshold value; and when the currently acquired image data is detected, excluding the target object in the invalid area. According to the embodiment, the detection accuracy of the target object can be improved.

The invention discloses an optical satellite remote sensing image cloud detection method. The method comprises the steps of counting a remote sensing image gray histogram and preprocessing the remote sensing image gray histogram; calculating a Gaussian mixture model component and an iteration maximum between-class variance threshold value according to the gray histogram, and further calculating a mixture threshold value; segmenting the remote sensing image by using the mixed threshold to obtain a cloud initial detection image; obtaining a contour boundary characteristic value and a cloud shadow matching characteristic value of a cloud area in the cloud initial inspection image, further removing the cloud area which is mistakenly detected, and obtaining a mistakenly overhauled positive image; performing missing detection thin cloud compensation on the cloud initial detection image to obtain a thin cloud compensation image; and carrying out comprehensive processing on the false overhaul positive image and the thin cloud compensation image to obtain a final cloud detection image. According to the method, the cloud area can be rapidly and accurately detected without manual or other auxiliary data participation, the cloud area which is mistakenly detected is removed in the detection process, the thin cloud area which is missed is compensated for, the cloud detection precision is improved, and the method is suitable for panchromatic remote sensing images and multispectral remote sensing images.

The invention relates to a manufacturing process of an FPC (Flexible Printed Circuit), in particular to a detection method for detecting whether flaws exist in drilling holes of an FPC semi-finished product. The detection method for detecting whether flaws exist in the drilling holes of the FPC semi-finished product comprises the following steps of: A, manufacturing of a checking base plate, to be specific, stacking a light semitransparent plate of which the color is different from that of the FPC semi-finished product when the FPC semi-finished product is subjected to drilling operation soon, enabling the light semitransparent plate and the FPC semi-finished product to commonly drill a hole and taking the light semitransparent plate as the checking base plate; and B, a checking mode, to be specific, firstly, placing the FPC semi-finished product to be detected on a detection platform, then stacking the checking base plate on the detected FPC semi-finished product for aligning and turning on a base lamp of the detection platform; secondly, observing the checking base plate; if light emergent colors of all holes in the checking base plate are uniform, indicating that no flaws exist in the drilling holes of the FPC semi-finished product; and if light emergent colors of a plurality of holes or hole edges or non-drilling positions of the checking base plate are different, indicating that the flaws exist in the drilling holes of the FPC semi-finished product. The detection method disclosed by the invention is used for detecting whether the flaws exist in the drilling holes of the FPC semi-finished product or not.

The invention discloses a pointer-type instrument identification method based on a patrol inspection robot, which comprises the steps of training an Adaboost classifier by utilizing instrument image data set and selecting an image centered by the instrument as a template image for each patrol inspection point; obtaining the pictures of the instruments to be detected at the designated patrol points, combining the Adaboost classifier and template images for rough positioning and accurate positioning, and screening out the dial area; according to the pointer color of the dial area, using different image preprocessing operations and pointer extraction operations to extract the pointer contour; according to the extracted pointer contour, calculating the direction of pointer rotation and the pointing reading by using the cosine theorem. The method of the invention utilizes machine learning, and can detect and identify the number of instrument representations under various illumination and robot attitude changes.

The invention discloses a medicine box detection method based on three-dimensional point cloud and image data fusion, and the method comprises the steps: inputting a medicine box image collected by a camera into an optimized U-shaped full convolutional neural network, and extracting a medicine box feature image; in the feature extraction part, a grouping residual error convolution module is used for extracting a preliminary feature image, a cavity space convolution pooling pyramid module is used for extracting feature image information of different scales of the preliminary feature image, and a mixed attention module is used for fusing the feature image information of different scales to obtain a two-dimensional fusion feature image; a segmented medicine box image is obtained through up-sampling; and whether the segmented medicine box image meets the detection requirement is judged, if not, three-dimensional information of the medicine box is extracted, a target in the image is positioned through a two-dimensional target detection network, and a cone point cloud corresponding to the two-dimensional detection frame is obtained according to a camera geometric imaging model. A Point Net point cloud network and a feature fusion network layer are adopted to carry out instance segmentation on the cone point cloud to obtain all target points. A target centroid is estimated by using a T-Net network, a target point cloud is moved to a centroid coordinate system, then estimation of parameters of a three-dimensional bounding box is obtained through a parameter estimation network and a feature fusion network layer, finally, the size and orientation of a medicine box are obtained, the type of the medicine box is judged, and the character 0 of the medicine box is recognized in combination with image information. The problems of long time consumption, high false detection rate and the like in existing medicine box detection are solved.

The invention relates to a lane line tracking method based on a Kalman filter. The method comprises the steps of initializing a counter; inputting lane lines obtained through Hough conversion detection at the moment k; brining Lane lines at k moments one by one with k- Matching the lane lines at the moment 1, and respectively calculating the deviation of the parameters at the two moments; updatingby matching with a counter; when the linear matching number is not smaller than the counter, starting a Kalman filter to track the lane line; otherwise, not tracking; and after 10 frames of lane lines are tracked, detecting and updating a straight line library again. According to the method, the problem that lane lines of one or more frames of images in a complex environment are missed or cannotbe detected is solved; a method for replacing a detection value with the estimation value is provided, the lane line detection stability is improved, when lane line detection of a plurality of framesof images fails, the estimation value can still be used for replacing the detection value for output, the lane line detection method is high in robustness and accurate in detection result, and the driving safety is improved.

The invention discloses a yarn hairiness detection device. The yarn hairiness detection device comprises a yarn conveying device, a yarn image acquisition device and a yarn image processing device, wherein the yarn image acquisition device comprises an image acquisition chamber and an area-array camera and a stroboscopic face light source which are arranged in the image acquisition chamber; an amplification lens is arranged in front of the area-array camera; the area-array camera and the stroboscopic face light source are oppositely arranged on the two sides of the image acquisition chamber; through holes are symmetrically formed in the other two sides of the image acquisition chamber; yarns penetrate through the through holes and are output by a power leather roller, and the conveying direction of the yarns in the image acquisition chamber is vertical to the amplification lens; and the area-array camera is in data connection with the yarn image processing device. By virtue of the yarn hairiness detection device disclosed by the invention, an image processing process can be simplified, and the yarn image acquisition efficiency and the hairiness detection efficiency are improved.

The invention discloses a subway public place people flow analysis method based on machine vision, and the method comprises the steps: S1, obtaining historical video flow data photographed by a subwaycamera, and extracting a subway people flow training data set; S2, loading a pre-training model of the YOLOv3 network, and initializing a network weight; S3, inputting the INRIA pedestrian data set and the subway pedestrian flow data training set into a YOLOv3 network for training to obtain a real target box label; S4, adding N times of up-sampling operation to obtain N feature maps with smallerscales, and changing the size of the input image; S5, performing network coarse training on the mixed data set, and optimizing the number, width and height of target boxes; S6, performing network finetraining on the subway people flow training data set; S7, detecting the subway people flow test data set by using the trained network model, and counting people flow; and evaluating the performance of the network model. The method is high in detection precision, the network is trained according to a strategy from coarse to fine. Boundary frame parameters are optimized, and the balance between thedetection speed and the precision is controlled.