35 results about "Foreground detection" patented technology

The invention provides a joint training and image processing method, device and system, and relates to the fields of artificial intelligence, optical imaging and the like. The method comprises the steps that a training sample is acquired, the training sample comprises a training image and a real label, and the real label is a foreground detection result corresponding to the training image; obtaining a modulation result after the training image is modulated by the first optical assembly, and determining an input image according to the modulation result; inputting the input image into the neural network model to obtain an output foreground detection result, and determining the foreground detection result as a prediction label; obtaining a foreground enhancement evaluation result corresponding to the input image, and determining comprehensive loss according to the prediction label, the real label and the foreground enhancement evaluation result; and updating structural parameters of the first optical component and / or model parameters of the neural network model according to the comprehensive loss. By applying the scheme, the task processing precision and the like can be improved.

The invention discloses a method for realizing foreground multi-modal detection based on adaptive dynamic weight fusion, which comprises the steps of multi-modal data acquisition, multi-modal data preprocessing, multi-modal feature extraction, multi-modal feature fusion, foreground detection and prediction result post-processing. The multi-modal feature fusion step comprises the following steps: a) respectively inputting the extracted RGB features and IR features into a weight prediction sub-network, wherein the sub-network generates a spatial weight map for each modal; each pixel value in the weight map represents the local confidence degree of the mode for foreground discrimination at the corresponding spatial position; b) normalizing the spatial weight maps corresponding to the modalities to enable the sum of the weights of the modalities at the same spatial position to be 1; and c) multiplying the normalized spatial weight map by the corresponding modal features element by element, and adding the modal weighted feature maps element by element to obtain a final comprehensive feature map for foreground detection. According to the method, the foreground detection precision and robustness are remarkably improved.

The invention relates to the technical field of computer vision and image processing, in particular to a foreground detection method fusing wavelet transform and a prompt mechanism, which is characterized in that in a dynamic background, high-speed motion and small target scene, F-measure is averagely improved by more than or equal to 0.30, and the detection precision is obviously superior to that of FFNet and ASCNet. The wavelet transform attention module suppresses background noise, highlights target edges and textures, and improves input feature purity. The CFF, MFR and PB of the decoder collaboratively fuse multi-scale features, redundant information is reduced, and target details and contour integrity are reserved. According to integral end-to-end training, single-frame reasoning only needs 28 ms, and real-time performance and accuracy are both considered. Therefore, the problems that in an existing neural network foreground target detection algorithm, input data are often influenced by noise and background interference, redundant information is easily introduced into a decoder in the multi-level feature fusion and transmission process, and consequently the detection result is inaccurate are solved.

The invention discloses an engine cylinder body and cylinder cover surface abnormal defect detection method based on a traditional algorithm, and relates to the field of engine part appearance defect detection. In order to solve the problem that in the prior art, abnormal defects cannot be effectively recognized and intercepted when defect samples are lacked or appear for the first time, the technical scheme provided by the invention is as follows: the method comprises the following steps: selecting a reference template image; constructing a background model based on the reference template image; carrying out registration and mask area extraction on the to-be-detected image to obtain a target area image, and carrying out noise removal and image smoothing on the target area image; based on the filtered image, extracting a first suspected defect area according to a preset threshold value; performing foreground detection on the filtered image based on a background model, and extracting an abnormal change region as a second suspected defect region; and obtaining a final defect detection result according to the first suspected defect area and the second suspected defect. The method is suitable for automatic surface abnormal defect detection and quality control work on an engine cylinder body and cylinder cover production line.

The invention discloses a sleeping posture recognition method and device and a storage medium, and the method comprises the steps: obtaining a current frame pressure image which comprises at least one target user; determining a foreground detection frame corresponding to a foreground area in the current frame pressure image; based on the current frame pressure image, predicting to obtain an initial sleeping posture recognition result of the target user; based on the foreground detection frame, correcting the initial sleeping posture recognition result; and smoothing the corrected initial sleeping posture recognition result to obtain a target sleeping posture recognition result of the target user. The current frame image is obtained, the foreground detection frame is determined, and the initial sleeping posture recognition result is corrected based on the foreground detection frame, so that the recognition accuracy is effectively improved. In the face of a complex scene, the foreground detection frame is used for comprehensive judgment, and interference factors such as noise, body movement and sleeping posture imaging missing can be better handled. And meanwhile, the corrected result is smoothed, so that the stability of the recognition result is improved.

The invention discloses an X-ray chip cavity detection method and system based on a VIBE algorithm. The method comprises the following steps: constructing a background model according to a cavity analysis algorithm; collecting a sequence image of a chip to be detected, and initializing the background model by using a front preset frame number of the sequence image; determining an adaptive segmentation threshold according to the image background complexity, and performing foreground detection and segmentation according to the adaptive segmentation threshold; adjusting the dynamic update rate of the background model according to the image background complexity, and updating the background model according to the dynamic update rate; performing morphological closing operation on the obtained foreground binary image; then connected component analysis is carried out, statistics is carried out on the total pixel number of all connected areas, the ratio of the total pixel number to the total pixel number of the chip to be detected is calculated, and the hole proportion is obtained; according to the invention, the background model of each pixel point can be established by analyzing the image of the chip area. And then, comparing each pixel of the current image with the background model so as to separate the foreground in the image.

This invention belongs to the field of computer vision and spacecraft on-orbit servicing technology, specifically involving a lightweight two-stage component segmentation method for non-cooperative space targets. Step one: Constructing an edge-guided foreground detection network, employing a MobileNetV3 backbone network combined with lightweight structures such as Ghost modules and depthwise separable convolutions, and introducing an edge-guided attention module to output a spacecraft foreground mask; Step two: Fusing the RGB image with this foreground mask to generate foreground enhancement features, constructing a fine-grained segmentation network within the foreground constraint region, and using the same lightweight backbone network to improve segmentation accuracy; Step three: Employing a staged training strategy, setting differentiated loss functions and learning rates, calculating the loss in the foreground region, and completing model training. This invention, through its lightweight architecture and staged segmentation design, compresses the model parameter count to 4.2M, achieving an optimal balance between segmentation accuracy and lightweight efficiency, enabling real-time segmentation of key spacecraft components on a spaceborne edge platform.

The application discloses a dynamic scene background modeling and foreground detection method based on a box plot, which comprises the following steps: saving the obtained video images into a background queue according to a dynamic frame extraction speed; based on the current background queue state, removing a moving interference target by using a box plot statistical method, modeling the background after removing the interference target, and obtaining a background model; and performing difference operation on each frame of the obtained video images by using the background model to obtain a foreground; and experiments prove that the dynamic scene background modeling method based on the box plot can effectively improve the modeling efficiency and controllability, and has important significance in the actual dynamic scene foreground detection task.

The invention relates to the technical field of image processing, in particular to a background difference-based discrete picture sequence foreground detection method, which comprises the following steps of: photographing and sampling an assembly line to obtain a picture sequence, establishing a background model and initializing; using the initialized background model to carry out front and back background coarse blocking and secondary coarse blocking on the picture sequence to obtain blocked pictures; according to the method, a relatively stable background is obtained through a fixed-interval fixed-point photographing mode, then a background model is established, it is ensured that a foreground target appearing in the moving background can be recognized and detected, and the efficiency and reliability of industrial automation are improved. Meanwhile, for factors such as sudden illumination change and object shielding in a pipeline environment, object movement speed variation and the like are detected, a background model is established through a discrete picture sequence, and foreground target extraction is carried out in cooperation with deep learning, so that the influence can be effectively reduced, and the detection effect and robustness of the method are remarkably improved.

This invention relates to a LOTO information-based and intelligent integrated security management method, comprising the following steps: 1) Optimizing the YOLO human body recognition algorithm under high-definition images, dividing the image into multiple regions with the same resolution as the standard image of the algorithm, and recognizing them separately; 2) Using OpenCV for image stitching, dividing, matching, and stitching these regional images so that the segmented object remains intact after the fusion of multiple images, avoiding duplicate recognition; 3) Adding motion foreground detection, using OpenCV for foreground detection; 4) Visual recognition of the number of locks; 5) Label recognition, using OCR to visually recognize security warning words; 6) Visual recognition of equipment status, developing algorithms for on-site equipment scenarios to identify the operating status of the equipment; This invention can reduce the risks caused by subjective factors of personnel in the traditional LOTO operation process, greatly improving LOTO security.

The application relates to the fields of computer vision and industrial intelligent detection technology, and discloses a foreground-background decoupling joint training method for an intelligent gangue selection target detection model. The method constructs an A background sample set, a B foreground sample set and a C real scene sample set; an initial target detection model containing a feature extraction layer, a feature fusion layer and a detection head is acquired, and initial weights are loaded; the A background sample set and the B foreground sample set are combined to form a foreground-background decoupling joint training set, and background sample identifiers or foreground sample identifiers are set for the training samples; in the first training process, a loss mask is generated based on sample type identifiers and target instance numbers, foreground detection loss and background suppression loss are respectively calculated, a weighted sum of the two is taken as a joint optimization target, and a foreground-background decoupling pre-training model is obtained; and then, taking the model as a starting point, scene fine-tuning is carried out on the C real scene sample set, and a target detection model is obtained.

This invention provides a method and apparatus for detecting abandoned objects. First, a background image of a detection area and multiple frames of images to be detected within the detection area are acquired. Each frame of the image to be detected and the background image are sequentially input into a foreground detection model for detection, resulting in the position region of each foreground object in each frame of the image to be detected. The position region is the area within the image to be detected where the bounding rectangle of each foreground object is located. For each foreground object, motion state detection is performed based on its position region to obtain the motion state detection result. Finally, for each foreground object, only if the detection result indicates that the foreground object remains stationary for a preset time is the foreground object determined to be an abandoned object. This method improves the detection accuracy of abandoned objects and avoids the influence of factors such as lighting and camera shake on detection accuracy.

The application discloses an image classification method and related devices and equipment, wherein the image classification method comprises: obtaining a to-be-classified image; performing feature extraction on the to-be-classified image to obtain a feature vector of the to-be-classified image; performing first classification processing, second classification processing, foreground detection and feature classification processing on the feature vector to obtain a first classification result, a second classification result, a foreground detection result and a feature classification result of the feature vector; and obtaining a classification result of the to-be-classified image based on the first classification result, the second classification result, the foreground detection result and the feature classification result; wherein the classification type of the second classification processing comprises the classification type of the first classification processing and the corresponding sub-type, and the classification type of the second classification processing is the same as the classification type of the feature classification. The above scheme can improve the image classification accuracy.

The application provides a port pollution ship monitoring method and device, the method comprises the following steps: acquiring a video image of a port area shot by a UAV, the video image is composed of continuous video frames; using a mixed Gaussian background model to perform foreground detection on each video frame pixel by pixel to obtain a foreground image containing a potential area; performing optical flow calculation on each foreground pixel point in the foreground image to determine the motion information of each foreground pixel point; according to the motion speed data and the motion direction information in the motion information, screening abnormal pixel points from the foreground pixel points that meet a preset abnormal condition; performing connected domain analysis on the abnormal pixel points, and determining a pollution area and a pollution type according to the analysis result; based on the detection time and the spatial position of each pollution area, combining ship automatic identification system data, calculating the correlation probability of each pollution area and a candidate ship to determine a pollution ship. The pollution ship is traced, and the port pollution is monitored in a wide coverage and all-weather mode.

The invention belongs to the field of safety early warning, and discloses a community safety event early warning method and system through AI analysis, and the method comprises the steps: obtaining a frame meeting a detection condition in a monitoring video: for a tth frame in the monitoring video, obtaining a set U1 of pixel points belonging to a foreground and a set U2 of pixel points belonging to a background in the tth frame through a foreground detection algorithm; judging whether the t-th frame meets a detection condition or not based on the U1; acquiring a set U3 of pixel points belonging to the foreground in the t-th frame based on an inter-frame difference algorithm, and updating a foreground detection algorithm based on U2 and U2; and carrying out AI analysis on the frame meeting the detection condition to obtain an analysis result. When the moving speed of the foreground object is slow, the foreground detection algorithm can still detect the foreground object with a high success rate.

A violent behavior detection method based on edge computing, constructing and training a deep learning model and a reinforcement learning method for violent behavior detection on a cloud server; a monitoring device performs foreground detection on a video frame, obtains a region-of-interest frame and uploads it to an edge server, and the edge server performs object detection, and feeds back the result of the region where people exist in the frame to the monitoring device; the monitoring device judges whether the number of people in the region where people exist exceeds a threshold, establishes a video frame buffer and calls a reinforcement learning method to screen key frames for the video frames, stores the key frames in the buffer, and if the buffer is full, uploads the video frames in the buffer as a group to the edge server, and the edge server calls the deep learning model to perform end-to-end inference on the group of video frames to obtain the probability of the existence of violent behavior in the group of video frames; the present invention can effectively reduce the consumption of computing resources and the occupation of network bandwidth in the entire process of violent behavior detection.

This invention belongs to the field of industrial anomaly detection based on computer vision, specifically involving a multi-task anomaly detection method based on target-oriented representation distillation and background denoising. The method includes constructing and training a multi-task anomaly detection model based on backdistillation, and using the trained multi-task anomaly detection network to detect anomalies in industrial product images. The multi-task anomaly detection model includes a teacher encoder, a student decoder, a single-class bottleneck embedding module, and a foreground detection module. This invention innovatively combines foreground detection with feature distillation, using target-oriented feature representation learning to solve the problem of mismatch between the pre-trained model and the distribution of industrial data.

The embodiment of the invention provides a method and equipment for detecting weeds in a fence and a medium. The method comprises the following steps: receiving a to-be-identified image and a shooting position which are sent by shooting equipment and are obtained by shooting a to-be-identified fence; obtaining a target historical image shot at a shooting position from a plurality of historical images obtained in advance, and performing feature alignment on the to-be-recognized image and the target historical image to obtain an aligned to-be-recognized image; performing foreground detection on the aligned to-be-recognized image by adopting a background model corresponding to the shooting position to obtain a foreground region; and carrying out weed detection on the foreground region by adopting a weed detection model to obtain a detection result. The method is used for achieving the effect of improving the detection accuracy and robustness of the weeds in the fence.

The invention is suitable for the field of detection, and discloses a multi-moving-target tracking counting method, device and equipment and a storage medium, and the method comprises the steps: obtaining a current frame of an input image sequence according to a time sequence, carrying out foreground detection and morphological operation, carrying out connected domain analysis, obtaining a plurality of connected domains, and carrying out the foreground detection and morphological operation; screening out a connected domain of which the area is greater than a preset connected domain area threshold as a current frame moving target; when the bounding box size of the current frame moving target is greater than a preset bounding box splitting threshold, splitting the current frame moving target into at least two current frame sub-targets; according to the method, the current frame moving target and the current frame sub-target which are not split are collectively called as the current frame effective target, the state of the current frame effective target is updated through association, creation and removal mechanisms, and after the state updating of all the current frame effective targets is completed, the number of the current frame effective targets is counted. And each sub-target can be identified and tracked independently.

The present invention discloses a method, device, and medium for rapid detection of dawn and dusk fog based on an adaptive threshold visual background extractor. The method comprises: obtaining H8 / AHI data and ground station fog observation marker data during the dawn and dusk periods and preprocessing them; constructing a time series detection dataset based on the H8 / AHI data; constructing an adaptive threshold visual background extractor based on the ViBe model, denoted as the ST-ViBe model; introducing an LBSP texture feature description operator when initializing the background model, and introducing a SILTP texture feature when establishing a pixel foreground detection parameter set; adaptively adjusting model parameters based on the input time series detection dataset; using the ST-ViBe model, obtaining preliminary fog detection results based on the time series detection dataset; and optimizing the preliminary fog detection results using traditional cloud removal and post-processing methods to remove residual clouds and afterimages. The present invention can accurately, efficiently, and rapidly detect dawn and dusk land fog.

The present disclosure provides an image stabilization method and device, electronic equipment and storage medium, relates to the field of artificial intelligence, in particular to the technical field of computer vision, intelligent transportation, vehicle-road cooperation and the like. The specific implementation scheme is: detecting the current image foreground of a perception device to obtain information of each obstacle in the current image; for each obstacle, determining a corresponding target stabilization frame in each preset stabilization frame of the current image; determining whether the obstacle shields the corresponding target stabilization frame based on the information of the obstacle; when not shielding, calculating the offset between the corresponding target stabilization frame and the corresponding stabilization frame in the reference image to obtain the stabilization offset of the corresponding target stabilization frame of the obstacle; when the stabilization offset of the corresponding target stabilization frame of at least one obstacle does not exceed a preset threshold, determining the new position of each obstacle based on the stabilization offset of the corresponding target stabilization frame of at least one obstacle and the information of each obstacle, thereby realizing the stabilization of the perception device.

A security discrimination method based on image recognition employs a high-definition camera to acquire real-time video streams of the target area. The real-time video stream is processed frame-by-frame into images to be detected. For each image, a Gaussian mixture model is used for background modeling and foreground detection, filtering out background interference and extracting the foreground target. The foreground target is then passed to a continuous frame detection module. If the detection result indicates no helmet is being worn, the next N consecutive frames of the image are detected. If all N frames show no helmet being worn and N is greater than or equal to a first preset value n, an alert is issued. The next M consecutive frames after the Nth frame are then detected. If all M frames show helmet being worn and M is greater than or equal to a second preset value m, the alert is terminated. Compared to traditional manual supervision, this invention significantly saves manpower and financial resources for real-time monitoring and also solves the problem of insufficient timely warnings from manual monitoring.

The invention discloses a micro-vibration monitoring method based on foreground detection and adaptive Euler video amplification, and belongs to the technical field of equipment micro-vibration monitoring, and the method comprises the steps: firstly collecting an electromechanical equipment operation video, extracting a foreground through a Gaussian mixture model to obtain a mask, and segmenting the mask into sub-regions through an SLIC algorithm; 4 vibration indexes of each sub-region are calculated and quantized intensity is fused, so that the number of layers of the pyramid and an amplification coefficient are matched; and building a Gaussian pyramid, a Laplacian pyramid and double IIR filtering to extract a target vibration signal, and reconstructing an image after amplifying and superposing an original signal according to parameters to obtain a vibration enhanced video. According to the micro-vibration monitoring method based on foreground detection and adaptive Euler video amplification, precise monitoring of micro-vibration of equipment is realized, adaptive matching of amplification parameters is realized, loss of details and artifacts are avoided, the identification degree of vibration signals is improved, meanwhile, the method adapts to a complex industrial environment, a clear vibration enhanced video is output, and the method is suitable for popularization and application. Faults of the assisting equipment can be found early, and operation and maintenance efficiency and safety are guaranteed.

This invention provides a method for detecting evidence left behind by a person and an integrated system for detecting such evidence. [Solution] The object detection method includes step A, which is a preprocessing step used for acquiring site images and foreground detection; step B, which is an object pairing and tracking step that detects objects in the original RGB images and selects objects that frequently appear in the monitored environment; and step C, which is an object detection step that detects whether an object is a common object at the site based on the object's overlap rate, displacement, and dwell time. This enables monitoring of whether objects at a site in a public space are artifacts.

The present invention provides a long-term anti-UAV visual tracking method based on spatio-temporal attention knowledge, including S1: obtaining an infrared video sequence of a UAV and extracting features from the video frame images; S2: constructing a tracking network model and training it, inputting the extracted feature images into the trained tracking network model, and outputting the predicted position and scale. The tracking network model constructs a two-stage re-detection architecture of cross-correlation RPN similarity learning RCNN based on a Siamese network, introduces a spatial position consistency function and a supervised foreground detector, and at the same time uses a triplet training mechanism to incorporate spatio-temporal attention knowledge into the learning process of the tracker, training the tracking network model, and realizing the position estimation and scale prediction of the UAV target in the infrared video sequence.

The present invention discloses a moving target detection method based on weighted low-rank and neighborhood group sparse decomposition, including: first, taking the grayscale image of each frame and normalizing the pixels to the interval [0, 1]; second, constructing an overlapping local image group for each video frame according to the circular neighborhood of the pixels and giving an expression of the neighborhood group sparse induced norm; third, decomposing D into a low-rank background matrix B, a structured sparse foreground S, and a Gaussian noise matrix E; fourth, constructing a weight W based on the structural information of S and the modeling result of the residual noise after background subtraction; fifth, establishing a final weighted low-rank and neighborhood group sparse decomposition model according to the above process; sixth, solving the model in the fifth step by the alternating direction multiplier method, updating the weight during the iteration process, and finally obtaining the foreground target. The present invention can detect foreground objects that are stationary for a short or long time, can handle complex dynamic backgrounds, and enhance the integrity of foreground detection, and is superior to the detection effects of other models in terms of the F-measure index and visual effects.

The application discloses a foreground object detection method and device based on a Gaussian mixture model, and an electronic device. Through application of the technical solution of the application, on one hand, a judgment condition of original model matching can be modified by calculating active points in an image. Thus, by performing model matching only on pixel points with a large change in pixel value in the image, consumption of meaningless model matching in a video stream image can be reduced. On the other hand, the result of twice Gaussian model matching can indirectly complete the effect of image denoising, and this way will not cause small targets to be filtered out. Thus, the problem of large foreground detection overhead in related technologies can be avoided.

The application is suitable for the field of detection, and discloses a multi-motion target tracking counting method, device, equipment and storage medium.The method comprises the following steps: acquiring a current frame of an input image sequence in time sequence, performing foreground detection and morphological operation, then performing connected domain analysis to obtain multiple connected domains, and screening out connected domains with an area greater than a preset connected domain area threshold as current frame motion targets; when the size of the boundary box of the current frame motion target is greater than a preset boundary box splitting threshold, the current frame motion target is split into at least two current frame sub-targets; the current frame motion target and the current frame sub-targets that are not split are collectively referred to as current frame effective targets, the state of the current frame effective targets is updated through an association, creation and removal mechanism, and after the state updating of all the current frame effective targets is completed, the number of the current frame effective targets is counted.The method can split oversized targets, and ensures that each sub-target can be independently recognized and tracked.

This application relates to the field of video analysis technology and discloses an algorithm for detecting highway litter based on image block segmentation, which successively includes: video frame extraction, automatic road area recognition, road segmentation, regional block expansion, expanded block grayscale image calculation, suspected moving target extraction, and litter discrimination. In this application, the detected picture area is segmented, and then each block area is processed independently. If there are obvious pixel changes in the expanded area corresponding to this block area, foreground detection is performed on this block area, and then the foreground targets obtained from each block area are merged to obtain the entire foreground image, and subsequent processing of litter target verification is performed on the foreground image. Due to the obvious spatio-temporal imbalance in highway traffic flow, there will be obvious scenes without vehicle flow continuously at different positions and different time periods in the monitoring area. By means of segmentation, the huge computational cost caused by detecting the entire picture can be avoided, and the computational amount can be effectively reduced.

The invention discloses a deep expansion foreground detection method fusing dictionary learning and incoherent terms, and the method comprises the steps: constructing a dual-channel dictionary learning mechanism, synchronously extracting the low-rank features of a background and the sparse structure features of a foreground, converting an additive decomposition model of a conventional RPCA into a mask product form through combining with a mask generation module, and carrying out the detection of the foreground through the mask product. Cooperative optimization of a foreground mask M and a low-rank background S is realized through end-to-end learning, a constraint layer of an incoherent item is introduced, and foreground and background separation is effectively promoted in a deep expansion network. According to the method, each part is solved through the ADMM, then each solved part is put into the network, the number of layers of the network directly corresponds to the number of optimized iterations, the structure is simple, the calculation efficiency is high, optimal parameter configuration is automatically learned through back propagation, and the requirement for manual parameter adjustment is reduced.