100results about How to "Improve accuracy and robustness" patented technology

The invention relates to a virtual-reality occlusion handling method based on depth image data flow. The virtual-reality occlusion handling method comprises three parts of construction of a scene point cloud model, three-dimensional space registration and virtual-reality occlusion handling and rendering. Firstly filtering and other processing operation are performed on depth data acquired by a depth camera and the normal vector of each point is calculated; then the camera attitude is calculated by using an iterative closest point algorithm according to the point cloud carrying the normal vector and the point cloud obtained by projection from a three-dimensional scene model through the last frame of camera attitude; then the point cloud of the current frame is fused into the three-dimensional scene point cloud model; when the scene is reconstructed, the color image feature points acquired by the depth camera are calculated in real time and three-dimensional space registration is performed by matching with the template image feature points; and then the space position relation and the occlusion relation of the virtual object and the three-dimensional scene are processed by the obtained camera attitude and rendered in real time. The method can be operated in real time on the present mainstream equipment, and the great virtual-reality occlusion effect can also be obtained when the resolution of the input data is low.

The present invention relates to a pedestrian detection method based on deep learning and multi-feature point fusion. The pedestrian detection method is characterized by at a training stage, firstly acquiring a pedestrian image under an application scene, marking the head and shoulder parts of the pedestrians in the image, and then using the pedestrian samples for the model training, wherein the model training comprises two steps of 1) taking the head and shoulder images of the pedestrians as the training samples, training a dichotomy model of the head and shoulder parts of the pedestrians; 2) using the model parameters obtained by the training in the step 1) to initialize partial parameters of a pedestrian detection model in a transfer learning manner. The pedestrian detection method of the present invention can overcome the problem that the pedestrians shield mutually to a certain extent, adopts a deep learning method to extract the pedestrian features, can better overcome the actual application problem that the factors, such as the pedestrian clothing, postures, backgrounds, illumination conditions, etc., change, also can effectively overcome the problems of the pedestrian multiple postures, the pedestrian multiple scales, the pedestrian mutual shielding, etc., and enables the pedestrian detection accuracy and robustness to be improved substantially.

The invention is applicable to the field of video and image identification, and provides a salient target detection method. A spatial-temporal cascaded neural network consists of a first fully convolutional network and a second fully convolutional network. The method comprises the following steps: inputting a current frame of image of a to-be-detected video to the first fully convolutional networkto obtain a spatial prior map; generating a temporal prior map according to the current frame of image and the optical flow map thereof; carrying out element operation on the spatial priori map and the temporal prior map to get a spatial-temporal prior map; and inputting the spatial-temporal priori map and the next frame of image to the second fully convolutional network to get a spatial-temporalsaliency map. According to the embodiments of the invention, in the detection of a salient target in a video with a complex scene, the spatial prior information of the video frame image and the temporal prior information based on the optical flow are integrated to eliminate a static salient region and generate a final spatial-temporal saliency map in a dynamic scene, so that more and richer information can be obtained in the dynamic scene, and the accuracy and robustness are improved.

The invention relates to the technical field of image processing, in particular to a target tracking method and a target tracking device. The method includes: acquiring a current frame image during target tracking, comparing the current frame image to a previous frame image to obtain a frame difference template; performing foreground detection by the current frame image and a pre-established background template to obtain a foreground template; fusing and processing the frame difference template and the foreground template to obtain a foreground image of a candidate area after fusion; and tracking a determined tracking target in the foreground image of the candidate area. Most background interference is eliminated from the foreground area of the extracted fused candidate target, so that the target can be tracked in the effective area, chances of getting trapped in local peaks during target tracking are slimmed, interference resistance of the target tracking method is further improved and algorithmic accuracy and robustness are improved.

The invention provides a gait identification method and a system thereof. The method is characterized by extracting a behavior characteristic, a position characteristic and a gait characteristic from acceleration data acquired from intelligent terminal carried by a user; using a pre-trained behavior identification model to identify a current user behavior according to the behavior characteristic; using a pre-trained position identification model to identify a current position of the intelligent terminal according to the position characteristic and the identified current user behavior; and using a pre-trained gait identification model to identify a user identity according to the gait characteristic, the identified current user behavior and the current position of the intelligent terminal. In the method, through a layered progressive identification mode, accuracy and robustness of gait identity identification are increased, a correlation sensor placing position and a direction do not need to be limited, and the method and the system are flexible and are convenient to use.

The invention discloses a target recognition and shape retrieval method based on hierarchical description. The method comprises the following steps of: extracting the profile feature of a target by a profile extracting algorithm, calculating a curvature value of each point on the profile target, extracting the angular point feature of the target by non-maximum value suppression, taking a profile segment corresponding to every two angular points as an overall feature describer of the target, carrying out hierarchical description on the profile points according to curvature, carrying out hierarchical description on the profile segments according to the importance degrees of value features, combining profile segments, the values of which are lower than evaluation thresholds, to form profile feature segments as partial feature describers of the target, carrying out normalization on the profile feature segments, and carrying out similarity measurement on the profile feature segments of different targets according to Shape Contexts distance. The method can be used for performing feature extraction on a target shape effectively, scale invariance, rotation invariance and translation invariance are achieved, the accuracy rate and the robustness in recognition are improved, and the computation complexity is reduced.

The invention relates to a dexterous hand tactile information based material classification method based on joint sparse coding, and belongs to the technical field of material classification. The method comprises the steps of: (1) collecting tactile information of objects serving as training samples; (2) classifying the training samples into i types according to different materials of the training samples, capturing each training sample, collecting tactile information to obtain a tactile time sequence, and establishing a training sample dataset; (3) extracting the features of the training samples according to the obtained training sample dataset, and establishing a tactile sequence dictionary phi (D); (4) capturing test sample objects required to be classified to obtain tactile time sequences of the test samples, and classifying the materials by the obtained tactile time sequence of each test sample to obtain the types of the test samples; and (5) performing the step (4) on all the test samples to obtain the material type of each test sample. According to the method, the tactile information based material classification is realized on the basis of the joint sparse coding method, and the robustness and accuracy of classification are improved.

The invention provides an AUV (Autonomous Underwater Vehicle) cooperative navigation method based on maximum correntropy unscented particle filter and belongs to the technical fields of nonlinear filtering and cooperative navigation. According to the method provided by the invention, a maximum correntropy unscented Kalman filter (MCUKF) algorithm is adopted, and state estimation in the AUV cooperative navigation process is solved. The method comprises the following steps: in the AUV cooperative navigation process, reconstructing a state equation and a measurement equation of cooperative navigation into a nonlinear recursive model, and processing by utilizing a maximum correntropy criterion; generating an importance probability density function needed in PF by adopting the MCUKF in a particle filter (PF) framework, and acquiring estimation of the AUV state according to an algorithm flow of the PF, so as to realize localization of the AUV and completing cooperative navigation. Accordingto the AUV cooperative navigation method disclosed by the invention, in AUV cooperative navigation in which measured noise has a burst value, more excellent performance than that of the conventional particle filtering, improved particle filtering and robust filtering can be achieved.

An image conversion-based heterogeneous image block matching method provided by the invention has the following steps: acquiring training samples and test samples, constructing an image conversion network, training the image conversion network, constructing a feature extraction and matching network, training the feature extraction and matching network, predicting a matching result. The invention overcomes the problem that the feature difference of the heterogeneous image extraction is large and inaccurate in the prior art, effectively reduces the matching difficulty, and improves the accuracyand robustness of the heterogeneous image block matching.

The invention relates to an attitude transformation data processing method and device, computer equipment and a storage medium, and relates to an artificial intelligence image processing technology. The method comprises the steps that a source image and a target three-dimensional posture are acquired, three-dimensional segmentation voxels including voxel category information are obtained based onsemantic segmentation reconstruction, the three-dimensional segmentation voxels are projected to obtain a corresponding target posture two-dimensional segmentation image, and objects in the target posture two-dimensional segmentation image are labeled based on the category information to obtain component categories; a target two-dimensional attitude corresponding to the target three-dimensional attitude is obtained, and the source image, the target attitude two-dimensional segmentation image and the features of the target two-dimensional attitude are extracted to synthesize an intermediate-scale transformation image; the source image, the three-dimensional segmentation voxel, the target two-dimensional attitude and the transformation image are cut to obtain component layer data of each object component, and component synthesis on the component layer data of each object component is performed to generate a component image; and the transformed image and the component image are fused to obtain a target attitude image, thereby improving the quality of the attitude transformed image.

The invention relates to a radar signal processing technology, discloses a gesture recognition method based on an MIMO millimeter wave radar, and solves the problems of low detection accuracy, limitedgesture recognition types and high cost in a traditional gesture recognition scheme. According to the invention, the MIMO millimeter wave radar is used for transmitting and receiving electromagneticwaves, after distance dimension Fourier transform is carried out on received sampling signals, primary AOA beam forming is adopted according to virtual array data of a distance prediction angle, constant false alarm rate detection is carried out on the formed signals to obtain distance azimuth angle and other information, a distance dimension signal corresponding to a target obtained in the constant false alarm rate detection is extracted according to a HeapMap signal generated by first beamforming, second two-dimensional AOA beamforming is performed, parameters such as a pitch angle, a Doppler speed and a signal-to-noise ratio are solved, and according to the obtained parameters, the position, the azimuth angle and the pitch angle are analyzed and accumulated in a multi-frame mode and compared with preset feature parameters of related gestures, and gesture recognition results are obtained and output.

The invention discloses an intelligent granary environment safety monitoring system based on a field bus. The system is composed of a granary environment parameter collection platform based on a CAN field bus, and a granary environment safety evaluation subsystem. The system realizes the intelligent detection of granary environment parameters and the intelligent evaluation of granary environment safety. Many problems still existing in the granary environment caused by the unreasonably designed and poor traditional granary environment multi-parameter detection equipment, the incomplete detection system and the like are solved. Based on the nonlinearity and large lag of granary environment parameter changes and the large area and complex structure of the granary environment, the defects of inaccuracy, low reliability and the like of a granary environment monitoring system are overcome, accurate detection and reliable classification of granary environment parameters are realized, and therefore, the accuracy and robustness of granary environment parameter detection are greatly improved.

The invention discloses a self-adaptive sliding mode control method and system and a mobile robot controller. According to the method, the advantages of self-adaptive control and sliding mode controlare integrated, and a sliding mode control law of which the switching gain can be adaptively adjusted is designed. The control law is composed of two parts, wherein the first part is the design of a sliding mode control law, the second part is the design of a switching gain self-adaptive updating law. The method is based on a kinematics model and a trajectory tracking error model of the mobile robot. A trajectory tracking error differential equation of the mobile robot is deduced, a sliding mode control law is designed by applying a constant-speed reaching law method based on a switching function, an adaptive updating law of switching gain is deduced according to the designed sliding mode control law, and the adaptive sliding mode control law can enable the mobile robot to quickly convergeto an expected trajectory and an expected speed.

The invention is applicable to the field of image recognition, A neural network model training method for ultrasonic displacement estimation is provided, includes inputting target block and search block into twin neural network, The two branches of the twin neural network perform feature extraction using DenseNet with shared weights, The characteristic graph is obtained, and the convolution calculation is carried out on the characteristic graph obtained from the two branches of the twin neural network, and the scores of the obtained cross-correlation numbers are compared with the real values,and the difference results are propagated backward, so as to realize the adjustment of the weights of each layer of the twin neural network and the optimization of the network. The embodiment of the invention utilizes the depth neural network to extract the deep semantic features of the ultrasonic radio frequency data, enhances the accuracy and robustness of the displacement estimation, establishes the connection between the currently processed data and the existing other data, fully utilizes and mines the features of the existing data, and improves the utilization rate of the data.

The present invention discloses a railway track semi-automatic detection method based on the integration of the reflection intensity and geometric features. The method comprises the steps of 1, respectively calculating the local geometric features of each laser footprint in laser scanning data, wherein the local geometric features are composed of the main direction and the normal vector of the neighborhood point distribution of the laser footprint; 2, conducting the point cloud clustering based on the reflection intensity and the local geometric features and extracting the point cloud of a railway track. According to the technical scheme of the invention, the semi-automatic extraction for the point cloud data of the railway track can be quickly and robustly achieved. Therefore, the extraction accuracy, the extraction efficiency and the automation degree of the point cloud data of the railroad track is improved. Moreover, the method is simple and easy to implement.

The invention relates to the field of vehicle identification and particularly relates to a method for identifying taxicabs in real time by utilizing video images. The method comprises the following steps of: establishing a sample library, and setting color category sets for the taxicabs; defining a color parameter boundary and an area threshold of each color category in the category sets, and training a support vector machine classifier; tracking the taxicabs to be identified in an input video frame; extracting the area of each color category in a tracking window, and comparing the area of each color category with the area threshold, wherein if the area of at least one color category in the tracking window is more than the area threshold, the tracking is finished, otherwise, the tracking continues; dividing the tracking window into a plurality of pixels moving in up, down, left and right directions, thus obtaining four sub-windows; and inputting the sub-windows and the standard tracking window into the support vector machine classifier respectively to obtain identification results, and performing polling statistics on the results. Strict experiments prove that the taxicab identification method can be used for improving the accuracy of judgment, and is high in judgment accuracy and robustness.

The invention discloses a real-time measurement method and device of a power of a microwave radiation source signal under interference of a non-stable broadband that can effectively suppress the impact of the non-stable broadband interference on measurement of the power of a narrowband signal. According to the method, after a radiation signal of a target radiation source is received, a calibration signal generated locally is coupled to a received signal, and then analog down conversion and data sampling are performed; next, while a power of a sub-band in which to-be-tested useful signals are is calculated, a power of a signal in a cancellation sub-band is calculated; an average power measurement value of the cancellation sub-band is subtracted from an average power measurement value of the sub-band of the to-be-tested useful signals so as to obtain a power measurement value of useful signals containing no interference and noise; meanwhile, an average power measurement value of the calibration signal is calculated in real time; then the power measurement value of useful signals containing no interference and noise and the average power measurement value of the calibration signal are compared with a power nominal value Pc of the calibration signal, so as to obtain a power measurement value of a calibrated microwave radiation source signal.

The invention discloses an optimization method for a line laser vision inertial system, and the method comprises the steps: S1, carrying out the preprocessing of measurement data; S2, initializing anestimator; S3, obtaining an initial value by adopting a loose coupling sensor fusion method; S4, repositioning the sliding window; S5, optimizing the global pose graph; and S6, according to the globalpose and the laser triangulation depth of each pose, fusing to form a final 3D model. According to the optimization method, the final 3D model can be optimized, and the accuracy and robustness of theline laser vision inertial system are improved.

The invention discloses a rapid generating method for a multitarget space signal. The method is characterized in that a signal source approaching real environment is provided for a multipoint locatingsystem to be verified by simulating the motion tracks of multiple targets, so that the accuracy and robustness of the verification result for the integral performance of the multipoint locating system are improved effectively. The rapid generating method has contribution to the further development of the multipoint locating technology.

The invention discloses an iris image quality detection method to solve the problem that a current method cannot be used for effectively judging the quality of a low-quality iris image sequence. The method comprises the following steps: obtaining an iris image sequence, calculating an overall definition index, and allocating quality weights; calculating an iridal local definition index; calculating an iridal availability index; calculating an iridal neighbourhood contrast ratio index; carrying out normalization processing on the overall definition index, the iridal local definition index, theiridal availability index and the iridal neighbourhood contrast ratio index of each frame image of the iris image sequence; calculating the combined quality score of the iris image sequence, and evaluating the availability of the iris image sequence. According to the iris image quality detection method, the image quality evaluation of complex environment imaging identified in remote-distance movement is realized.

The invention provides a license plate detection and recognition method based on multi-video frame information and a nuclear phase light filtering algorithm. Different deep learning models are respectively constructed to carry out license plate area detection and license plate information identification of a single-frame image, and multi-frame information tracking fusion is carried out by using akernel-based correlation filtering algorithm, so that the accuracy of license plate identification and positioning can be improved, the calculation efficiency is relatively high, and the method can beused for real-time processing.

The invention provides a Parkinson's disease finger tapping action recognition method and system, a storage medium and a terminal. The Parkinson's disease finger tapping action recognition method comprises the following steps: acquiring video data including hand tapping actions of a Parkinson's disease patient; obtaining hand skeleton sequence data in the video data based on an OpenPose algorithm;constructing skeleton sequence features based on the hand skeleton sequence data; constructing a deep learning model for multi-stream fine-grained skeleton sequence action recognition based on the skeleton sequence features; and according to the deep learning model, identifying a to-be-identified finger tapping action. According to the Parkinson's disease finger tapping action recognition methodand system, the storage medium and the terminal, recognition of the Parkinson's disease finger tapping action is achieved based on the hand posture estimation algorithm and the deep learning algorithm, the accuracy is high, and the practicability is high.

In order to improve the accuracy of target tracking, the invention provides a feature detection method based on sparse coding, which belongs to the technical field of target tracking in the field of computer vision. The invention comprises the following steps: S1, extracting the local feature points by using the FAST corner detection algorithm based on the local threshold, calculating the local gradient direction of the feature points, sampling an image block centered on the feature points as a training sample; S2, grouping the dictionary elements according to the local gradient direction of the feature points to obtain an over-complete dictionary; S3, using the obtained dictionary to carry out sparse representation of the test sample, and then dividing the image block into blocks to construct sparse features, and realizing target tracking according to the detector of the sparse features. The invention improves the accuracy and robustness of target tracking by utilizing the sparse characteristics of the sparse encoding learning target. The dictionary elements of different groups are trained according to the local gradient direction of the image block to reflect the local directioninformation of the image block and to improve the accuracy of target tracking.

The invention discloses a cloth image retrieval method based on a convolutional neural network, and the method comprises the steps: carrying out preprocessing of a textile fabric image, zooming of theimage through bilinear interpolation, and carrying out the normalization and other preprocessing operations; designing a convolutional neural network as a classifier; training the neural network by using a classified loss function and gradient back propagation iteration to obtain a feature extractor; performing feature extraction on the retrieval graph and the fabric library to obtain a 1024-dimensional feature vector; and calculating the similarity of the two feature vectors by adopting an L2 measurement method, and sorting to realize recognition of textile fabric image retrieval. Accordingto the invention, contour spatial position feature extraction can be carried out on the target shape, and recognition of the target with occlusion is realized. The method has scale invariance, rotation invariance and translation invariance, so that the problem of incomplete contour recognition is effectively solved, and the accuracy and robustness of target recognition and shape retrieval are improved.

The invention provides an identity authentication method and device applied to the field of artificial intelligence AI. The invention can be specifically applied to intelligent terminal equipment suchas an intelligent robot for user identity authentication, and comprises the steps of acquiring information of multi-modal biological characteristics of a to-be-recognized user wherein the multi-modalbiological characteristics comprise at least two biological characteristics of the to-be-recognized user; performing identity authentication on the to-be-identified user in parallel according to at least two biological characteristics; and determining an identity authentication result of the to-be-identified user according to an identification result obtained by performing identity authenticationon the to-be-identified user in parallel, with the identity authentication result being obtained based on a confidence coefficient of matching of the multi-modal biological feature and a preset biological feature. According to the technical scheme, interference factors existing in identity authentication through a single biological feature can be effectively avoided, and therefore the robustnessand accuracy of the identity authentication method are improved.

The invention relates to a marine target association system and method based on high-low orbit optical satellite observation. The marine target association system comprises a target detection and tracking module, a target detection and motion feature extraction module and a target multi-level multi-feature association module. The target detecting and tracking module is used for detecting and tracking a marine target by utilizing the high-orbit optical satellite image sequence to obtain motion trail information of the marine target. The target detection and motion feature extraction module is used for performing marine target detection and motion feature extraction by utilizing the low-orbit image to obtain the position and heading information of the target. The target multi-level and multi-feature association module is used for carrying out marine target multi-level and multi-feature association based on the motion trail information obtained by the target detection and tracking moduleand the position and heading information obtained by the target detection and motion feature extraction module. According to the method, the monitoring capability of the marine target can be remarkably improved, and meanwhile, the method has very high target association accuracy under a complex observation background and is easy to implement.

The invention discloses a method for diagnosing and self-repairing soft and hard faults suitable for health monitoring. The method forms a routing link, the first node of the link is a cluster head, and the second node is a backup cluster head. It is proposed to use M.A‑B .The method proposed by Abdo extracts two characteristic parameters for node fault detection and structural damage identification respectively from the original collected data, so as to avoid the interference of faulty nodes on structural damage. The self-healing wireless sensor network structure system plans to adopt a hybrid hierarchical topology. In addition to the original sensor nodes, it is considered to set up monitoring nodes in the network to monitor the network status. The method applies the cooperative communication mechanism to the multi-stream transmission problem of the wireless network to realize the maximum throughput of the network and ensure the stability and reliability of the network.

The invention discloses a microblog emotion classifier establishing method and device. The microblog emotion classifier establishing method comprises the following steps that multiple microblog data is obtain; word segmentation is conducted on microblog texts to obtain a word set; word-word emotion association relation and word-emotion association relation are obtained according to a large number of unannotated microblog texts; artificial annotating is conducted on the preset number of microblog texts to endow emotion labels so as to obtain an annotated microblog emotion data set; the word-word emotion association relation, the word-emotion association relation and the annotated microblog emotion data set are integrated to obtain multi-source heterogeneous emotion information, and a mathematical model is established to obtain a microblog emotion classifier. By the adoption of the method, the microblog emotion classifier can be obtained by fusing the multi-source heterogeneous emotion information, accordingly the emotion tendency expressed by microblog messages are classified, the robustness and accuracy rate of the classifier are improved, the cost is low, and the method is simple and easy to achieve.

The invention discloses a synthetic aperture sonar motion compensation method for reducing the accumulated error of a DPC (Displaced Phase Center) method, which comprises the steps of receiving echo signals by synthetic aperture sonar, performing aperture division on an azimuth synthetic aperture within a complete synthetic aperture length to obtain a plurality of sub-apertures, wherein each sub-aperture contains a plurality of echo signals; performing motion error estimation and compensation on inter-ping echo signals by adopting the DPC method in the sub-apertures; and estimating and compensating a motion error between the sub-apertures in a correlation manner according to sub-aperture image data between the sub-apertures, wherein the step further comprises the sub-steps of solving matching points of sub-aperture images, performing error estimation and compensation based on the matching points and compensating the estimated error into the original echo data or using the estimated error for correcting an imaging grid.

The invention discloses an intelligent scoring positioning method based on deep learning, which takes the positioning of oral calculation questions as an example, and comprises the following steps: shooting a plurality of oral calculation question pictures by using a mobile phone, and dividing the pictures into a training set and a test set; Utilizing a labelImg marking tool, and marking the position of a mouth calculation question on each picture in the training set by using Bouncing Box; Generating an xml file, and converting the xml file into a txt file; modifying YOLOv3 algorithm, adding the category of the oral calculation question of the pupil in the classification, and training the data set; After training is completed, saving the weight, testing the pictures in the test set, obtaining the Bounding Box of each oral calculation question in each picture through regression, and the oral calculation questions of the pupils are positioned. According to the invention, the positioningfunction of the plurality of calculation questions in the picture can be realized, and the accuracy of the test result is relatively high, so that the workload of manual paper inspection can be reduced.