240results about How to "Good robustness" patented technology

The invention discloses a deep learning-based image high-density population counting method. The method comprises the following steps of S1, establishing a depth complementation convolutional neural network by utilizing a deep learning framework caffe; S2, performing image data enhancement on an image according to operations of angle rotation, image multi-scale zooming, image mirroring and image pyramid zooming; S3, performing Gaussian kernel fussy normalization processing on the enhanced image data to obtain a real crowd density graph, outputting an estimated density graph and the real density graph by the network, and performing continuous iterative training optimization on the whole network structure according to a loss function; and S4, inputting a crowd picture and a tag picture to the network for training, and performing continuous iterative optimization to obtain a trained network model finally. According to the method, the end-to-end convolutional neural network is designed; a picture is given and input, and the estimated density graph corresponding to the picture is output, so that an estimated crowd number is obtained; and by outputting the density graph, more useful information is reserved.

The invention provides a method for controlling the position of a self-adaptive robust of a motor servo system based on an extended state observer. The method comprises the steps of 1, establishing a mathematic model of the motor servo system; 2, configuring an adaptive law to estimate the uncertain parameters in the motor servo system; 3, configuring the extended state observer to estimate the uncertainty of the motor servo system; 4, configuring a position controller of the self-adaptive robust of the motor servo system based on the extended state observer; 5, determining relevant parameters and functions in the motor servo system to enable the position output of the motor servo system to accurately and progressively track the expected position command, and enabling the input of the motor servo system to not have the chattering phenomenon. The invention also relates to a system for controlling the position of the self-adaptive robust of the motor servo system based on the extended state observer.

The invention discloses a digital watermark embedding and extraction method based on chaos sequences. The digital watermark of the chaos sequences is used for carrying out scrambling, and the pertinence of the original watermark and the scrambled digital watermark is removed through the scrambling, so the digital watermark has the features like flat noise. Thereby, the transparency of the digital watermark is improved. The position of the watermark embedding into a DCT block is determined through the chaos sequences, the safety of the algorithm is improved, and better robustness on the large-area cutting attack can also be realized. In the watermark embedding process, firstly, image sub blocks carry out discrete cosine transform (DCT), and the energy can be converted into low-frequency factor compression blocks; and after the conversation, the obtained low-frequency component matrix A carries out singular value decomposition (SVD), and at the same time, the position of the watermark embedding into the DCT low-frequency compression blocks can be determined and the embedding intensity can be regulated through the chaos sequences. Thereby, the calculation complicity can be reduced, in addition, the embedding quantity is also increased, the robustness of the watermark is improved, and the capability is improved for resisting ordinary attack, so the invention has wider practicability.

The invention discloses a deep-learning and cloud service-based face identification attendance system and method. The system comprises a face detection module, a data wireless transmission module, a cloud server and an attendance information management webpage, wherein the face detection module is connected with the data wireless transmission module; the data wireless transmission module is connected with the cloud server via network; a deep-learning network training module is established in the cloud server, face images are pre-trained, and feature vectors are saved; face images are extracted and transmitted to the cloud by virtue of the face detection module and the data wireless transmission module, and is input into the deep-learning network training module as a testing sample to match with faces, the match result is saved to a database, and the attendance information management webpage is interacted with the database to acquire attendance information. According to the system and method, matching identification is prevented from being performed in mass data by classified storage and classified calling, so that storage efficiency and matching identification efficiency can be improved, and the system has better robustness and higher efficiency.

The invention discloses a bionic dry adhesion mechanism based on loading of a composite constant-force spring. The bionic dry adhesion mechanism is composed of an adhesion sole unit, a composite constant-force spring loading mechanism, an electric motor, a transmission mechanism and a mechanism frame, wherein the adhesion sole unit is used to touch and adhere an target object; the composite constant-force spring loading mechanism is used to control adhesion / releasing of an adhesion sole; and the electric motor and the transmission mechanism are used to drive loading / unloading of the composite constant-force spring loading mechanism. The mechanism frame is used to support the whole mechanism and keep a relative position of each part of the whole mechanism. According to the invention, a bionic dry adhesion material is adopted, and adhesion force is not influenced by the shape of the target; the composite constant-force spring is used, so that the adhesion unit is applicable to adhesion surfaces with various materials, and the mechanism has robustness aiming at local adhesion failures; and through cooperative operations with equipment such as a mechanical hand, target capture efficiency can be greatly increased.

The invention discloses a power quality disturbance classification method based on a sparse automatic coded depth neural network. The method comprises the following steps of preprocessing a power quality disturbance original data set, inputting a preprocessed data sample into a stacked sparse automatic encoder to carry out automatic feature learning to obtain a feature expression, a step of transmitting an extracted feature expression to a softmax classifier for fine adjustment and further adjusting network parameters to obtain a trained classification model, and a step of inputting the a testsample into the classification model for power quality disturbance classification. According to the method, nine power quality disturbance signals including two kinds of composite disturbances can beaccurately identified, and the method has good robustness.

The invention discloses a mechanical scanning meter wave radar based angle measurement method which mainly aims at solving the problem that the angle measurement of a mechanical scanning meter wave radar is low in accuracy by the traditional single pulse method. The mechanical scanning meter wave radar based method comprises step 1, dividing an antenna into two sub-matrixes equally and transmitting pulse signals; step 2, performing coherent accumulation on received echo signals and obtaining data after the accumulation; step 3, performing DET (Discrete Fourier Transformation) calculation on the data after the accumulation; step 4, finding out a point in a doppler channel, reserving M points on the left side and the right side of the point and setting rest points to be 0, wherein the point is corresponding to the doppler frequency; step 5, performing IDFT (Inverse Discrete Fourier Transform) calculation on the data which are set to be 0; step 6, obtaining a sum beam and a difference beam according to obtained two groups of data after the IDFT calculation; step 7, performing the sum and difference beam single pulse angle measurement on the sum beam and the difference beam to obtain an off-axis angle of a target; step 8, adding the off-axis angle to a reference angle to obtain the accurate angle of the target. According to the mechanical scanning meter wave radar based method for improving the single pulse angle measurement, the accuracy of the angle measurement is high, a plurality of targets can be distinguished, and the mechanical scanning meter wave radar based method can be applied to the target accurate positioning and multi-target detection of the mechanical scanning meter wave radar.

The invention relates to an extended range electric vehicle energy management method on the basis of fuzzy control, which particularly comprises the following steps that: 1, a vehicle control unit acquires a storage battery SOC (State of Charge) and bus demand power data; 2, the vehicle control unit judges whether the storage battery SOC is less than or equal to 90 percent, the step 3 is executed if yes, and if no, a range extender is controlled to be switched off by a CAN (Controller Area Network) bus; 3, a fuzzy control module carries out fuzzification on the storage battery SOC and the bus demand power data according to a membership function; 4, fuzzy reasoning is carried out on the fuzzified data according to a set fuzzy rule and the membership function; 5, defuzzification is carried out on a reasoning result by utilizing a centroid method and output power distribution values of the range extender and a storage battery are output; and 6, the vehicle control unit sends the output power distribution values of the range extender and the storage battery to the range extender and the storage battery by the CAN bus. Compared with the prior art, the extended range electric vehicle energy management method has the advantages of strong adaptability, capability of promoting the whole vehicle performance and the like.

An embodiment of the invention provides a take-out delivery time prediction method and a take-out delivery time prediction device. The take-out delivery time prediction method comprises the steps of:acquiring order information of a take-out order, and determining historical influencing factors and current influencing factors corresponding to the take-out order according to the order information,wherein the historical influencing factors are determined based on historical take-out delivery data; and determining delivery duration of the take-out order according to the historical influencing factors and the current influencing factors of the take-out order and a nonlinear prediction model. Since the historical influencing factors and the current influencing factors influencing the take-outdelivery duration are fully considered when predicting the delivery duration of the take-out order, the take-out delivery duration can be predicted more comprehensively. Since a gradient boosting treemodel and a random forest model are combined to construct the nonlinear prediction model used for predicting delivery duration, the influence of nonlinear influencing factors on the take-out deliveryduration are fully considered, thus the nonlinear prediction model has better fitting effect, better robustness and better delivery duration prediction effect when compared with the traditional linear regression model.

The invention provides a flexible satellite attitude maneuvering control method based on tracking of a time-energy consumption optimal trajectory, and relates to a maneuvering control method for a flexible satellite attitude. In order to solve the problem of contradiction between bias of rotational inertia and loss time and achieve time-energy consumption optimal control, under the condition that the influence of flexible vibration is taken into consideration, an optimal angle tracking trajectory and an optimal angular speed tracking trajectory corresponding to the optimal angle tracking trajectory are calculated in real time from the maneuvering start moment according to a time-energy consumption optimal control method, it is guaranteed that the angle optimal trajectory calculated by attitude angle tracking of a rolling channel has short loss time and meanwhile has good robustness to bias of rotational inertia through PD control, and optimality of time and energy consumption of a fly wheel are taken into consideration at the same time. The flexible satellite attitude maneuvering control method is suitable for maneuvering control over the flexible satellite attitude.

The invention discloses an airplane angle-of-attack protection control method. The airplane angle-of-attack protection control method comprises the steps of 1 collecting an airplane signal; 2 calculating a maximum angle-of-attack; 3 calculating an obtained angle-of-attack threshold; 4 calculating a driving rod vertical alarm displacement and an instruction angle-of-attack; 5 obtaining a differential branch signal, a proportion branch signal, an integration branch signal and an integration branch logic operation result; 6 outputting, controlling and stabilizing the larger one of the integration branch signal and the integration branch signal; 7 determining the value of step 6; 8 determining a gain so as to adjust a real angle-of-attack of the airplane; and 9 obtaining a yield value of the differential branch signal, the proportion branch signal and the integration branch signal when the airplane is in different states. Through adoption of the airplane angle-of-attack protection control method, a connectional initial lead of the angle-of-attack protection function can be reduced; the angle-of-attack protection function can be realized; and a good robustness is provided to the system.

The invention relates to a data fusion method and a data fusion device for a low-cost integrated navigation system. The data fusion method comprises the main steps of: obtaining a course psiI of a carrier by carrying out strapdown inertial navigation calculation on output data of a microinertia measuring unit in the integrated navigation system; obtaining a course psim of the carrier by carrying out calculation on output data of a magnetoresistive sensor in the integrated navigation system; obtaining a course psiG of the carrier by carrying out calculation on output data of a satellite receiver in the integrated navigation system; and carrying out superposition on the psiI, the psim and the psiG by certain weight according to the set optimum estimated performance index of the course of the carrier, and obtaining an optimum estimation value of the course of the carrier. The embodiment of the invention has the advantages that under the confinement of the optimum performance index with minimum estimated variance, an estimation coefficient can be adjusted automatically according to the dynamic characteristic of the carrier, the optimum course estimation in the moving process of the carrier is given, the self adaptability is stronger, the robustness is good, and the practical application in engineering is convenient.

The invention discloses a deep wavelet neural network-based polarimetric SAR (Synthetic Aperture Radar) image classification method, and aims to mainly solve the problem of classification accuracy reduction caused by fewer characteristics or unreasonable characteristic extraction in the prior art. The method is implemented by the following steps: inputting an image; performing preprocessing; selecting samples; training a deep wavelet neural network by utilizing a training sample; extracting characteristics; performing classification; calculating classification accuracy. According to the method, the deep wavelet neural network is trained layer by layer, so that the problem of gradient diffusion in case of more network layers is solved; moreover, high-dimensional characteristics capable of reflecting essential properties of data, describing detail characteristics of the data and highlighting differences between different ground object types can be extracted; the deep high-dimensional characteristics of the data are extracted by virtue of the deep wavelet neural network, so that the problem of fewer characteristics or incomplete and unreasonable characteristic learning in a classification technology is successfully solved, and the classification accuracy of a polarimetric SAR image is improved.

The invention discloses a color image search method based on scale invariant feature transform (sift) seed region growing. The color image search method based on the sift seed region growing comprises the following steps: (1) choosing a sift seed point for an image to be detected; (2) dividing the image in to a plurality of regions by the adoption of a region growing method; (3) combining the regions according to the rule; (4) aiming at each region to extract a region characters; (5) calculating the similarity between regions of images in an image library and the region of the image to be detected; (6) structuring a weight bipartite graph; calculating the similarity between the image to be detected and the images in the image library; (7) arranging the images in the image library in descending order according to the similarity of the image to be detected. By the adoption of the mode, the color image search method based on the sift seed region growing can reduce the error problem due to inaccuracy of the images division and make up the defect that a single character search image is used.

The invention discloses a road semiautomatic extraction method based on wavelet detection and ridge line tracking, which mainly solves the problem that the prior road extraction method can not quickly and accurately extract curved roads. The method comprises the following steps: firstly, carrying out region growing and smoothing on high-brightness interference regions; then carrying out wavelet edge detection on the region to be extracted to obtain an edge detection graph; adding a center point correction method in the extraction process; in the edge detection graph, calculating the road azimuth according to a ridge line tracking algorithm and a minimum variance estimate and gradiometer algorithm; and finally, gradually extracting the road along the road azimuth. The invention has the advantages of high speed, accuracy and strong robustness, can obtain favorable extraction effect especially for roads with severe curvature change, and can be used for map updating, image matching, target detection and digital mapping automation.

The invention discloses a method for vehicle positioning and region segmenting in an image. The method includes the steps of obtaining an original image from collecting equipment; positioning a vehicle image area in the original image by means of Harris corner points and horizontal edge line segments, and obtaining a positioned vehicle image; carrying out color salient region extraction, smooth salient region extraction, horizontal edge sparse region extraction and horizontal edge dense region extraction of the positioned vehicle image; positioning a typical area of a vehicle according to extraction results and prior knowledge, and therefore segmenting an engine cover area, a vehicle face area and a vehicle window area. According to the method, the vehicle and other vehicles can be positioned at the same time, efficiency is high, accuracy is more than 95%, during area segmentation of the vehicle, interference of factors such as reflective effects of a windshield, nonuniform lighting and pollution can be effective eliminated, and robustness is good. The method can be widely used in the field of image processing.

The invention discloses a pitch detection method which researches aiming at the problem of a poor pitch detection technology in content-based music retrieval. The pitch detection method comprises the following steps of: converting a music signal to a frequency domain by virtue of Fourier transform to calculate, carrying out the first step of pitch detection on the signal according to a harmonic peak value method to find five low-frequency harmonic peaks, carrying out ascending sort according to the values of frequencies, then calculating the ratio among the frequencies, determining a group of pitch candidate sequences according to the data measured by an experiment, then carrying out pitch detection on the original music signal by a cepstrum method, combining the pitch sequences obtained by the two methods into a new pitch candidate sequence, finally finding a pitch corresponding to the lowest cost, that is, the standard pitch obtained by the method, by virtue of a confidence degree and viterbi optimal algorithm. The method disclosed by the invention is great in robustness and good in anti-noise performance.

The invention relates to a target tracking method based on ORB characteristics point matching. The target tracking method comprises: adopting an ORB (Oriented FAST and Rotated BRIEF) characteristic point set to represent a target object, and in a tracking process of the target object, iteratively updating the ORB characteristic point set by continuously adding new characteristic points and cutting outliers. A nearest neighborhood search algorithm is used between two continuous frames to find out matched characteristic point pairs, and then a motion transformation model of the tracking target is worked out according to the characteristic point pairs by using method that combines random sampling accordance with a multiple transformation model. The target tracking method based on ORB characteristics point matching is better in robustness and higher in efficiency, when being used for processing scenes with complex appearance change, disorder background, illumination variation and shielded target object.

The invention relates to a Rochester model-naive Bayesian model-based data classification system, which comprises a data processing module, a sampling module, a modeling module and a data testing module, wherein the data processing module divides an original sample set into a saturated layer and a lacking layer according to the input missing value ratio of each sample variable in the original sample set and relativity among the sample variables and sample attributes; the sampling module randomly extracts a training sample variable and a testing sample variable from the saturated layer and the lacking layer to form a training sample set and a testing sample set of which each comprises the saturated layer and the lacking layer respectively; the modeling module models training samples in the saturated layer through a Rochester regression model and models the training samples in the lacking layer through a naive Bayesian model to obtain a hybrid dynamic model with the Rochester regression model and the naive Bayesian model; and the data testing module inputs testing samples in the saturated layer into the Rochester regression model in the hybrid dynamic model, inputs the testing samples in the lacking layer into the naive Bayesian model in the hybrid dynamic model and performs a test to obtain and output scoring results. The Rochester model-naive Bayesian model-based data classification system is integrated with the functions of the Rochester regression model and the naive Bayesian model so as to have complementary advantages and can be widely applied to the financial industry, retailing and the telecommunication industry.

The invention discloses a mount element detection method based on color image segmentation and gradient projection positioning, and belongs to the technical field of automatic optical detection of mount elements. The mount element detection method based on color image segmentation and gradient projection positioning is characterized in that a designed image acquisition system is used to acquire color images; three-color (red, green and blue) images are converted into an HSI color model which is formed by three parameters (H (Hue), S (saturation) and I (Intensity)); as the HSI model directly uses the hue and the saturation having no relation with brightness when the HSI color model extracts the color information, the mount element detection method based on color image segmentation and gradient projection positioning is accurate and highly efficient in segmentation; one the above basis, positioning and detection of elements are carried out; positioning mainly uses the color and geometrical characteristics, and a positioning method based on gradient projection and a principal component characteristic value detection method are proposed; and the mount element detection method based oncolor image segmentation and gradient projection positioning has the advantages of high positioning accuracy and efficiency and accurate detection.

The invention discloses a method for classifying and identifying a neonatal pain expression and a non-pain expression based on sparse representation. A training sample is used to construct a redundant dictionary. A test sample is taken as a linear combination of the training sample in the redundant dictionary. The specific sparse representation is used to carry out pain and non-pain expression classification and identification. Expressions of smile, a quiet state, crying without pain, a mild pain state and an intense pain state can be effectively identified. The method has good robustness for noise of a neonatal face image and a shielding problem. The invention provides a new method and a new approach for developing a neonatal pain automation assessment system.

The measurement of geometric data of scene and bodies in the scene via single frame of image includes the following steps: taking image of the scene and inputting it into computer; extracting the data of point and straight line in reference plane of the scene, parallel lines perpendicular to the plane and the reference distance; finding out the mapping matrix between the reference plane and the image plane; calculating the blanking point in the direction perpendicular to the reference plane and the blanking line of the reference plane; extracting the boundary information of the bodies to be measured in the image, sectional fitting to form the straight line or quadric curve and finding out the coordinate of center point of bodies; and measuring geometric data, including spatial coordinates, distance, angle, area, volume, etc. of scene and bodies in the scene.

The invention discloses a block plane fitting-based image binaryzation method, comprising: dividing an image into many blocks; calculating the binaryzation threshold of each block; obtaining a relation expression about the binaryzation threshold of the whole image and the location of each pixel point in the image by fitting the binaryzation threshold of each block; and conducting image binaryzation by using the relation expression. The invention also discloses a device corresponding to the method. The invention has the prominent advantages that both local and global binaryzation information is considered, good robustness is achieved both for images with uniform illumination and images with non-uniform illumination, the threshold can be accurately obtained even when there are few blocks, and accuracy and efficiency are balanced.

The invention discloses an occluded pedestrian re-identification method based on massed learning and deep network learning. According to the method, multiple types of occluded training samples are generated from original non-occluded training samples through an occlusion simulator, the generated occluded training samples and the original training samples form a united training set used for model training, meanwhile, occluded and non-occluded classification losses are added into pedestrian classification losses, and a multi-task loss function is used to replace a single-task loss function in the past. In this way, pedestrian re-identification under occlusion is effectively coped with, and priori information of occlusion and non-occlusion is considered to perform feature extraction during deep network feature learning. Experiments indicate that through the method, the performance of an existing deep network in occluded pedestrian re-identification can be substantially improved, and the method has high application value.

The invention discloses a one-dimensional distance multi-classifier fusion recognition method based on class confidence. The method comprises the steps that K nearest neighbor samples of a preprocessed actual measurement one-dimensional distance image signal sample amplitude feature are extracted by the Euclidean distance measure; the reliability evaluation of the neighborhood sample class is acquired through each sub-classifier classification result to acquire the evaluation matrix of each classifier; and each class confidence of the sample is acquired through the evaluation matrix to realize classification. According to the invention, the method is based on the decision fusion theory; the K-nearest neighbor idea is combined; the nearest neighbor and the Bayesian criterion are used to acquire the class confidence of the target sample to complete target classification; and compared with a single classifier, a traditional voting method and other decision fusion methods, the method has the advantages of high recognition accuracy and great robustness, and is a practical one-dimensional distance image decision fusion recognition method.

The invention discloses an intelligent indication method of a pointer instrument and a device thereof. The method comprises the following steps: obtaining an image to be processed from a video streaming, carrying out image median filtering processing and image enhancement processing on the image to be processed to obtain a primary-processing image; on the basis of an Otsu threshold value segmentation method, carrying out binarization processing on the primary-processing image to obtain a binarization image, carrying out Laplace transform on the binarization image, carrying out xor processing on the binarization image and an image obtained via the Laplace transform, and carrying out the median filtering processing on the image obtained by the xor processing to obtain a secondary processing image; carrying out refining processing on the secondary processing image, carrying out linear detection processing on the image obtained by the refining processing to identify a pointer straight line of the pointer instrument; and according to the parameter information of the pointer instrument and the identified pointer straight line, calculating a scale value where the pointer of the pointer instrument points to at the moment. The pointer identification robustness of the pointer instrument can be improved.

The invention relates to a sensor node locating method and device based on sequences. The sensor node locating method comprises the following steps that a sensor node to be located collects the signal strength indication values of a plurality of beacon nodes within the communication range; regional division is conducted according to the coordinate position of each beacon node and the communication range of each beacon node, and sorting is conducted on the signal strength in each region to obtain the signature sequences; a plurality of detection sequences and the signature sequences undergo matching calculation sequentially; the sensor node to be detected collects the coordinate positions of beacon nodes beyond a first jump range and the coordinate positions of a located quasi beacon node; a perpendicular bisector is respectively and sequentially drawn between each beacon node beyond the first jump range and a corresponding beacon node within the first jump range and between each quasi beacon node and a corresponding beacon node within the first jump range to obtain a final region to be located; the centroid of the region to be located serves as the coordinate position of the sensor node to be located. The sensor node locating method and device based on the sequences have robustness, have higher matching precision and lower the requirements for the intensity of the beacon nodes.

The invention discloses an OFDM (Orthogonal Frequency Division Multiplexing) frequency synchronizing method based on a Zadoff-Chu sequence, belongs to the technical field of wireless communication, and aims at overcoming the shortcoming that the precision of the traditional frequency synchronizing algorithm is limited by the timed synchronizing accuracy and complicated to reach. The method is characterized in that the fractional frequency offset can be estimated through two repetitive sequences, the integer frequency offset can be estimated according to the influence of the frequency deviation on the self-correlated influence of the Zadoff-Chu sequence, wherein the estimation can cover the whole OFDM bandwidth, all operations are finished at the time domain, and any time-frequency transform is saved. Compared with the traditional frequency offset estimation algorithm, the method has the advantages that the operation complexity is small, and the robustness is high; the verification shows that the method has high performances under a high-speed moving and multi-path environment.