400results about How to "Improve object detection performance" patented technology

The invention relates to a radar target detection method based on dual-channel convolutional neural network false alarm controllability, and belongs to the technical field of radar signal processing.The method comprises the following steps of firstly, preprocessing radar echo signals, and constructing a training data set by using signal time-frequency information and amplitude information; then,constructing a dual-channel convolutional neural network model which comprises a dual-channel feature extraction network, a feature fusion network and a false alarm controllable classifier, and inputting a training data set to carry out iterative optimization training on the dual-channel convolutional neural network model to obtain an optimal network parameter and a judgment threshold; and finally, preprocessing a real-time radar echo signal, and inputting the preprocessed real-time radar echo signal into the trained dual-channel convolutional neural network model for testing to complete target detection. The method is suitable for radar target detection in a complex environment. According to the method, radar signal multi-dimensional features are intelligently extracted and fused. The detection performance is improved. The false alarm rate control is achieved. The actual demands of radar target detection are met.

A system, method and program product for providing a video surveillance system that enhances object detection by utilizing feedback from a tracking system to an object detection system. A system is provided that includes: a moving object detection system for detecting moving objects in a video input; an object tracking system for tracking a detected moving object in successive time instants; and a tracker feedback system for feeding tracking information from the object tracking system to the moving object detection system to enhance object detection.

The invention discloses a pulse Doppler radar system based on multiple detection signals and a signal processing method. The pulse Doppler radar system comprises a plurality of transmitters, wherein each transmitter respectively transmits orthogonal pulse detection signals synchronously at different pulse repetition frequencies (PRF), and the PRF of different detection signals meet stagger relationship, the receiving system of airborne radar is composed of a receiver or a plurality of receivers. The signal processing method processes the echo corresponding to different transmission signals via target detection, to obtain respective target detection results. The interference accumulation time of the pulse Doppler radar can be significantly prolonged under single PRF, thereby improving target detection performance. The invention utilizes the stagger relationship among different PRF, to process ambiguity resolution treatment on the detection result having distance and Doppler ambiguity under different PRF, to obtain correct target distance and Doppler information.

The invention discloses a video target detection method based on a convolutional gating recurrent neural unit, and solves the problems of tedious steps and low detection precision in the prior art byusing video data time sequence context information. The method comprises the implementation steps of data set processing and network pre-training. The method comprises steps of selecting a reference frame, and estimating a reference frame feature based on the current frame feature; carrying out time sequence context feature learning based on the convolutional gated recurrent neural unit; performing weighted fusion on the time sequence related characteristics; extracting a target candidate box; carrying out target classification and position regression; training to obtain a video target detection network model; and verifying model effects. According to the invention, by introducing a characteristic propagation mode of a current frame estimation reference frame, and establishing a time sequence relation between the current frame and reference frame characteristics, the current frame is enabled to have reference frame information by using the convolutional gated recurrent neural unit, andthe feature quality of the current frame is enhanced by using a weighted fusion mode. And under the condition of low time cost, the detection precision is improved, the complexity is reduced, and themethod can be used for video target detection.

The invention provides a model compression method and device, electronic equipment and a computer storage medium. The method comprises the steps that training sample data are acquired, and the training sample data comprise label sample data; Respectively training the teacher network model and the student network model by using the training sample data to obtain an adaptive distillation loss function and a focus loss function; and performing back propagation on the student network model according to the adaptive distillation loss function and the focus loss function to obtain a trained studentnetwork.

The invention discloses an improved algorithm based on parameter estimation and compensation of a quadratic phase function, and belongs to the technical field of signal and information processing. Thealgorithm performs range walk correction on slow time-range frequency domain echo signal considering range walk and Doppler diffusion through Keystone transform, and then eliminates a coupling relation between the range frequency f and the slow time tn in a second-order phase item corresponding to the acceleration in a narrow-band condition; meanwhile, a folding factor corresponding to the blindspeed is searched, a folding factor compensation item is constructed to correct the range walk caused by the blind speed; then the target acceleration is estimated by using the quadratic phase function, and a second-order phase item corresponding to the acceleration is constructed to compensate Doppler diffusion of the echo signal; and finally, long-term phase-coherent accumulation is performed soas to realize high-speed weak target detection of bistatic radar. The method disclosed by the invention is high in operation speed, stable in performance and applicable to detection of weak targets with a low signal-to-noise ratio and difficultly estimated target parameters.

The invention relates to the field of robot vision, pattern recognition and machine learning, in particular to an underwater robot target detection method, a system and related equipment, aiming at improving the robustness of target detection technology to underwater target occlusion, deformation and illumination changes. The object detection method of the invention comprises the following steps:obtaining an original image to be detected; normalizing the pixel value of the original image to be detected, and obtaining the image to be detected after preprocessing; the preprocessed image being input into the target detection network for detection, and the bounding frame of the region of interest and the probability of belonging to each target class being obtained; according to the bounding box of ROI and the probability of belonging to each target class, the improved non-maximum suppression algorithm being used to obtain the bounding box and the class of the target object, wherein, a deformable convolution neural network is used to extract feature map in the target detection network, and the candidate region method is used to detect the target. The detection method of the invention improves the detection precision under the condition of guaranteeing the speed.

The invention belongs to the technical field of target detection, and provides a radar target detection method based on combined adaptive normalized matched filter, wherein the method is used for detecting weak movement targets under the sea clutter background. The method includes the specific steps that a coherent pulse string transmitted by a beam position resided by radar receives echo data of all distance units to form a distance-pulse plural matrix; partition and combination are conducted on the echo data matrix along the pulse dimension according to a certain rule; the noise wave covariance matrix estimation values of the distance units to be detected are computed for the distance units to be detected of echo data blocks obtained through partition; the values of detection statistic of all Doppler channels are solved through the estimation values and the data of the distance units to be detected, the maximum value of the detection statistic in each echo data block is obtained, and the values of all combination detection statistic are solved through the product of the maximum values; the judgment threshold is worked out through a Monte Carlo method according to the set false alarm probability and the number of the combined echo data blocks; if the value of the combined detection statistic is larger than the judgment threshold, it is judged that the distance units to be detected have a target, and if not, it is judged that the distance units to be detected do not have a target.

The invention discloses a self-adapting space interference suppression method of a one-dimensional phase scanning three-coordinate radar, which comprises the steps that a radar system distributes detection pulse resources, a signal processing subsystem detects and receives echoes of antenna of all rows, azimuth angles and pitching angles, carries out digital down conversion processing, weights wave beams at equal intervals in directions of the pitching angles to obtain space-distance domain data, then carries out one-dimensional fourier transformation, extracts interference feature information in a frequency domain for interference judgment, and records the azimuth angle and the pitching angle of interference when judging the interference; and at the same time, the signal processing system carries out wave beam weight water-flowing solving for multi-beam weighting to form multiple received wave beams reset in an interference direction by a phase-only method according to an interference angle recorded in the last scanning period and an expected direction of the current target. The method applies self-adapting space interference suppression to a pulse doppler radar according to characteristics of the one-dimensional phase scanning three-coordinate radar, and survivability and target detection capability of the radar in an interference environment are effectively improved.

A learning apparatus comprises a plurality of detection units configured to detect a part or whole of a target object in an image and output a plurality of detection results; an estimation unit configured to estimate a state of the target object based on at least one of the plurality of detection results; a classification unit configured to classify the image into a plurality of groups based on the state of the target object; and a weight calculation unit configured to calculate weight information on each of the plurality of detection units for each of the groups based on the detection results.

The invention discloses a suppression device for the distance side lobe of a discontinuous spectrum high-frequency radar based on Bayes regularization and a suppression method thereof., which is characterized in that the suppression device comprises a transmitting antenna, a receiving antenna, an antenna switch, a matched filter, a distance side lobe suppression system, a frequency synthetic system and a frequency spectrum monitoring system; the suppression method thereof comprises the following steps: monitoring the frequency spectrum of outside electromagnet environment on a real-time basis and selecting a plurality of frequency points transmitted from the frequency spectrum; carrying out frequency synthesis to a discontinuous frequency value selected by the frequency spectrum monitoring system and transmitting; carrying out matching treatment to transmitting signals and receiving signals after the reflection echo of the transmitting signals is received; constructing the cross response of the transmitter and the receiver of each distance unit; and constructing regular functions according to a target spectrum model and suppressing the distance side lobe by adopting the interative method. The invention has the advantages of good suppression effect of the distance side lobe of the discontinuous spectrum high-frequency radar, insensitive Doppler, good suppression effect of noise, and improved target detection performance of the radar.

The invention discloses an RNA binding protein prediction method based on a multi-scale attention convolutional neural network. The RNA binding protein prediction method comprises a training stage anda prediction stage. The training stage comprises the following steps: preprocessing RNA data, encoding the RNA data, constructing a neural network and training network parameters. A mathematically abstract statistical mode of RNA is converted into a matrix form, the matrix form is input into a pre-designed attention mechanism-based multi-scale convolutional neural network, and parameters in the neural network are trained by using an Adam optimization method by minimizing a designed special cross entropy loss function. In the prediction stage, RNA sequence data with four basic groups as basicunits are input into the network, the probability of whether binding sites corresponding to binding proteins exist in the RNA data is output by the last layer of the neural network, and therefore theprediction result of the RNA sequence category is obtained. The method can improve the prediction precision.

A multi-radar system is configured such that radars A and B perform synchronization so that their transmission timings and frequency bands are substantially the same. For this reason, in a case where the radars A and B operate as bistatic radars, the radars A and B yield the same detection results from reflected waves from a place on a surface of a target T where the detection object regions of the radars A and B overlap. Therefore, the signal-to-noise ratio can be improved by synthesizing the target detection results obtained by the radars A and B operating as bistatic radars. This results in improved target detection performance.

The invention discloses an unmanned ship navigation collision avoidance radar detecting method. The method comprises the following steps that omnibearing target searching is carried out on an observed sea area through a radar antenna and a transceiver, reflection echoes of a target are received, and target original video signals with clutter are obtained after amplification frequency conversion detection; the original video signals are sent to a signal collecting module through a signal interface module for A / D conversion, and target original digital video signals with clutter are obtained; software-oriented trace point extraction, signal processing, clutter rejection and target detecting tracking data processing are automatically carried out on the target original digital video signals through a ruggedized computer, meanwhile, signals of a gyrocompass, a log, a GPS, an AIS and an optoelectronic device of the ship are received, moving platform motion compensation and target fusion are carried out, various on-sea targets are captured and tracked, a target track is set up, and then tracking detection is carried out. Comprehensive detection, tracking and recognition of the unmanned ship in the complex sea condition on multiple batches of targets on the sea are achieved.

The invention discloses a frequency diversity MIMO radar distance-angle decoupling beam forming method, and mainly solves problems of impossibility of forming distance-angle decoupling self-adaptive responses of an existing frequency diversity array. The realization steps are that 1, the frequency of emission signals of the frequency diversity array is designed; 2, the emission signals and receiving end echo signals of the frequency diversity array are obtained; 3, vector outputting is carried out on the receiving end echo signals to obtain snapshot vectors of the echo signals; and 4, self-adaptive beam forming is carried out on the snapshot vectors to obtain a distance-angle two-dimensional beam forming directional diagram. According to the invention, the degree of freedom of emission of the frequency diversity array is fully utilized to form a controllable degree of freedom in a dimension of distances and angles on the basis of a frequency diversity MIMO radar system, and two-dimensional beam forming of distances and angles is realized through self-adaptive beam forming technology. The frequency diversity MIMO radar distance-angle decoupling beam forming method can be used for combined detection of distances and angles, and suppress interference signals related to the distances.

The invention discloses a passive time reversion reverberation inhibition method based on forward predication, comprising: first, processing submarine reverberation and target echo via modeling and multiplying window operations, obtaining a signal sub space corresponding to the target distance, using forward predication to evaluate a submarine reverberation sub space corresponding to the target distance, deriving a group of optimal weighted vectors w of SRA, using the optimal weighted vectors w to process weighted sum on the echo signal x(r, omega) received by SRA, thereby restraining the reverberation component of the echo signal and improving the target echo signal component of the echo signal most, to further improve the echo signal mixed ratio. The invention can effectively restrain reverberation, keep or enhance target echo, to effectively improve echo signal mixed ratio, improve the target detection property of active sonar and improve practicality.

The invention belongs to the technical field of radar signal processing, relates to localized suppression processing of radar clutters and discloses a method for partial combination clutter suppression in airborne MIMO radar three-dimensional beam space. The method comprises the steps that firstly, an expression of a space-time data vector x is obtained and a space-time two-dimensional guide vector of a target is established; secondly, an array element-pulse domain of the space-time data vector x is converted into a three-dimensional beam domain through three-dimensional discrete Fourier transform; thirdly, a space-time data vector z generated after dimensionality reduction and a space-time two-dimensional guide vector, generated after dimensionality reduction, of the target are obtained; fourthly, according to a cost function of dimensionality reduction space-time self-adaptive processing, a weight vector of a dimensionality reduction processor is calculated; fifthly, space-time data processed by clutter suppression are obtained through the weight vector of the dimensionality reduction processor. The method can solve the problems of the large operand and large sample requirements and is applied to the scene of radar clutter processing.

An infrared image background suppression method which is based on Wavelet and Curvelet transforms relates to an infrared image processing method, and the method is used for solving the problem that the method for suppressing background interference in the infrared image processing field also suppresses the useful target information while realizing the background clutter suppression. The Wavelet transform decomposition and the Curvelet transform decomposition are respectively carried out on an original infrared image; high-frequency decomposition coefficients under different resolutions are respectively retained for the Wavelet coefficient and the Curvelet coefficient, the low-frequency decomposition coefficient is set as zero; image reconstruction is carried out on the decomposed original infrared image by the retained Wavelet coefficient and the Curvelet coefficient respectively; image fusion is respectively carried out on a Wavelet coefficient reconstruction image and a Curvelet coefficient reconstruction image which are obtained by step 4 by a multi-resolution decomposition frame, thereby obtaining the final infrared image background suppressing result.

The invention discloses an unmanned aerial vehicle video small-object detecting method based on super-resolution reconstruction. The unmanned aerial vehicle video small-object detecting method comprises the steps of selecting one input image as a reference frame, and selecting three consecutive image frames for performing sub-pixel displacement estimation on the reference frame; then placing the displacement estimation result of the four image frames into a high-resolution image grid; and estimating pixels which are lost in the high-resolution image grid, thereby obtaining an objective image with a relatively high resolution. Afterwards, an objective template is extracted from the objective image, and the characteristic of the objective image is calculated. Then the reconstructed objective image is divided for obtaining a plurality of objective area blocks. Characteristic extraction and characteristic identification are successively performed on all objective area blocks, thereby finishing preliminary detection for the object. Afterwards, false object elimination is performed, thereby obtaining a final detection result.

An object detection apparatus that detects an object in a vicinity of a vehicle includes: (a) an image processing circuit configured to: (i) derive vectors representing movement of feature points in captured images acquired periodically by a camera that captures images of the vicinity of the vehicle; and (ii) detect the object based on the vectors; and (b) a controller configured to (i) acquire a velocity of the vehicle; and (ii) set a parameter that affects a number of the feature points based on the velocity of the vehicle.

A system for reducing CFAR loss due to sea clutter is disclosed. The system includes a first channel tuned for CFAR gain in a spatially correlated background and a second channel tuned for low CFAR loss in spatially uncorrelated backgrounds. Each of the channels employs a distribution free CFAR using rank ordered statistics to establish a constant false alarm rate. The output of each channel is fused by a hit correlation function and the stream of combined hits is processed by a target centroiding function.

The invention provides a real-time target detection method based on the oriented gradient two-value mode (ORBP) and soft cascade SVM, and aims at solving the problems that the target detection in the prior art is low in instantaneity and robustness. The method comprises the steps that 1) characteristic of the oriented gradient two-value mode is described; 2) a soft cascade classifier SVM is established; 3) characteristic training is carried out on the soft cascade classifier SVM; and 4) a target window is tracked and updated. ORBP has the advantages that rotation, scale, translation and brightness are not changed, the soft cascade SVM improves the target detection robustness in complex scenes, and tracking of the target window improves the instantaneity of target detection. The method provided by the invention can be applied to man-machine interaction and intelligent traffic monitoring fields, and the target detection performance is excellent.

The invention relates to the field of target detection based on a frequency-modulated continuous wave radar, and particularly belongs to a multi-target number detection method and device based on thefrequency-modulated continuous wave radar. The method comprises the steps that the frequency-modulated continuous wave radar transmits a linear frequency-modulated signal, and the received echo signaland the transmitted linear frequency-modulated signal are mixed; the mixing signal are filtered and discretized to obtain a discrete intermediate frequency signal containing distance and speed information of a plurality of to-be-measured targets; windowing processing is performed on the discrete intermediate frequency signal; two-dimensional fast Fourier transform is performed on the windowed discrete intermediate frequency signal to obtain two-dimensional amplitude spectrum information; a two-dimensional combined adaptive constant false alarm rate algorithm is adopted to perform preliminaryestimation on the number of to-be-measured targets, and a two-dimensional spectrum peak is updated; and based on the two-dimensional spectrum peak, final estimation is performed on the number of the to-be-measured targets by adopting multi-scatter target condensation processing, thereby determining the number of the targets. According to the invention, the detection performance is effectively improved under the condition of relatively low time cost.

The invention discloses a monocular perception correction method and device based on sparse point cloud, and a storage medium. The method comprises the steps of collecting the original camera data ofa monocular camera and the original sparse point cloud data of a radar sensor; processing the original camera data to obtain a three-dimensional detection result of a plurality of targets in an imageplane, the three-dimensional detection result comprising a target depth value and a two-dimensional bounding box; obtaining a conversion matrix; mapping the original sparse point cloud data to a corresponding position of an image plane based on the conversion matrix to obtain a point cloud projection depth map, setting a point cloud frame for each two-dimensional boundary frame in the point cloudprojection depth map, the point cloud projection depth map comprising a plurality of projection points corresponding to the original sparse point cloud data, and each projection point comprising a point cloud depth value; and correcting target depth values of a plurality of targets based on the point cloud depth values of the projection points contained in all the point cloud frames. The method isadvantaged in that by designing the point cloud frame characteristics, accuracy of correcting the target depth value is improved.

The invention relates to a weak object detection method on the basis of generalized S transformation, which pertains to the field of image processing, in particular to a weak object detection method by applying the two-dimensional generalized S transformation of an image. The method carries out two-dimensional generalized S transformation to the original image I (x', y') at first to obtain a generalized S transformation result S (x, y, kx, ky); then detection is carried out to a weak object in the image according to the generalized S transformation result S (x, y, kx, ky); by fixing the value of spatial frequency kx or ky, a four-dimensional S transformation result S (x, y, kx, ky) is reduced into three-dimensional data; and the combination of frequency component power of a vertical section image (y, ky) or (x, kx) corresponding to each value selected along x or y direction is compared with a threshold value so as to determine the positions of row and line of the weak object in the original image. The method carries out generalized S transformation to the original image; on basis of the generalized S transformation result, by dimension reduction and visualization processing, the information of the S transformation domain is utilized to detect the weak object, thus being capable of effectively overcoming the influences by obstruction factors such as noise, flooding wave, cloud, and the like, in the spatial domain.

The invention relates to an SRDF (Space-Range-Doppler Focus) moving object accumulation detection method for a frequency diverse array radar, and belongs to the technical field of radar signal processing and detection. The method comprises the steps: carrying out the echo demodulation and matched filtering of the frequency diverse array radar, constructing a distance-azimuth two-dimensional vectorof a receiving array element signal; realizing distance / azimuth joint estimation based on a spatial spectrum estimation algorithm; implementing the high-resolution sparse domain Doppler extraction based on sparse fractional Fourier transform to complete the space-frequency focusing processing of a moving object; constructing a space-range frequency domain detection unit map, and detecting the moving object. The method fully utilizes the flexible freedom degree and energy aggregation of the frequency diverse array radar in space, distance and Doppler, achieves the coherent projection and SRDFin a multi-dimensional space, integrates the pulse compression, angle estimation and Doppler filtering of the conventional radar signal processing, is high in accumulation gain and parameter estimation accuracy, and can improve the radar moving target detection capability in complex environments.

The invention discloses a fish finder based underwater fish target detection and recognition research method. The method comprises the following steps: firstly performing classification training on different fish underwater sonar images, and obtaining an effective recognition classifier; secondly, using a fish finder to acquire sonar images and performing preprocessing; recombining the solar images into a three-channel image through an initial background model of a self-organizing neural network background subtraction model, transferring the three-channel image to an HSV color space, using Gaussian weights to initialize the background model, and calculating the minimum distance between a pixel and a pixel value in the background model corresponding to the pixel; performing determination, classifying the pixel into the background model if the minimum distance is less than a threshold, and updating the model; performing shadow determination if the minimum distance is greater than or equal to the threshold; determining that the pixel is a foreground if the pixel is not the background or the shadow, sending a foreground target to the trained effective recognition classifier to performrecognition and classification, and counting the number, the type and the size of the fish. The invention can quickly and accurately detect the fish target, and can effectively and correctly identifyand classify the fish target, and can be applied to the monitoring of the underwater fish group.

The invention discloses a detection method for spot target on a satellite, the method comprises: a step of radiation correction and geometric rectification processing of image data, a step of acquiring target image data by background suppression processing of the image data, a step of acquiring target oversampling image data by eliminating particle peaks in the target image data, and a step of acquiring target position data by utilizing the target oversampling image data. The detection method of the invention assists in realization of detection of image data and acquirement of target on a satellite.

The invention discloses a method and a device for realizing augmented reality based on deep learning and a mobile terminal, and belongs to the technical field of computer application. The method comprises the following steps: separating a planar image from a depth image collected by a depth camera; Carrying out deep learning on the planar image through a pre-trained neural network model; determining a target in the planar image, extracting a pixel coordinate and a corresponding depth value of each pixel of the target in the depth image; calculating a three-dimensional coordinate of the targetaccording to the pixel coordinate and the depth value; Therefore, the augmented reality is realized, the realization of the augmented reality gets rid of the limitation of an application scene, the traditional augmented reality technology is upgraded to an augmented reality method based on deep learning, the application scene of the augmented reality is greatly expanded, and the target detection capability of the augmented reality is enhanced.