23874results about How to "Small amount of calculation" patented technology

The invention relates to a flight path planning method based on a sparse A* algorithm and a genetic algorithm and belongs to the technical field of flight path planning of an unmanned aerial vehicle (UAV). According to the characteristics of flight path planning, the method comprises the following steps: planning an initial reference flight path by the utilization of the sparse A*search (SAS) algorithm, wherein constraint conditions are combined into algorithm search, so that useless nodes in a search space can be effectively deleted and the search time is shortened; and when emergent threat occurs during real-time flight of the UAV, performing real-time flight path planning by the utilization of the genetic algorithm and generating a flight path with local optimum or approximate optimum until the threat disappears and the UAV returns the original global optimum reference flight path and continues flying. The method provided by the invention has high real-time performance and rapidity; the searched flight path is closer to the actual UAV optimal flight path; and the method can be applied to the technical fields of robot path planning, urban vehicle path planning and the like under complex environments.

The invention relates to a movement human abnormal behavior identification method based on template matching, which mainly comprises the steps of: video image acquisition and behavior characteristic extraction. The movement human abnormal behavior identification method is a mode identification technology based on statistical learning of samples. The movement of a human is analyzed and comprehended by using a computer vision technology, the behavior identification is directly carried out based on geometric calculation of a movement region and recording and alarming are carried out; the Gaussian filtering denoising and the neighborhood denoising are combined for realizing the denoising, thereby improving the independent analysis property and the intelligent monitoring capacity of an intelligent monitoring system, achieving higher identification accuracy for abnormal behaviors, effectively removing the complex background and the noise of a vision acquired image, and improving the efficiency and the robustness of the detection algorithm. The invention has simple modeling, simple algorithm and accurate detection, can be widely applied to occasions of banks, museums and the like, and is also helpful to improve the safety monitoring level of public occasions.

The invention discloses a face detection and recognition method based on skin color segmentation and template matching. A face is segmented with the use of a face skin mixing Gauss model according to the clustering characteristics of face skin color in an YCbCr color space and a region of which the color is close to face skin color is separated from an image so as to achieve the purpose of quickly detecting the outer face of a face. A light compensation technology is used to overcome the influence of brightness on face detection and recognition. Adaptive template matching is used to overcome the influence of a skin-color-alike background on face detection and recognition. A secondary matching algorithm is adopted to reduce the amount of computation during the matching process, and singular value features of a face image are extracted to recognize a face and achieve the purpose of reducing the number of dimensions of a characteristic value matrix in the process of face feature extraction. The face detection and recognition method can quickly detect a face, improves the accuracy of face detection and recognition, is strong in practicability and has strong popularization and application values.

The invention discloses a target following method in a complex interference scene, comprising the steps of: detecting a moving target by confirming the initial position of every moving target which appears in video, detecting with self-adapting Gaussian background models and separating a moving target area with a morphologic treating method, modeling for target feature by acquiring a plurality ofmatching feature messages for each detected moving target area, fusing a plurality of feature messages and building a target feature template, searching the matching area by searching the area which is matched with the target feature template in a specific area in current video as a target following location of current video, updating the target feature template by updating the feature of the target feature template of current video to finish the self-adapting update for the target feature template after locating the target according to the feature in the current video target area and the feature of the target feature template in the former three videos.

The invention provides a steel plate surface defect detection method based on multistage characteristics of a convolutional neural network and relates to the technical field of steel plate defect detection. The method comprises the following steps: selecting a baseline network, pre-training the baseline network, and establishing a special defect detection data set for fine-tuning training; building an overall detection network and a multistage characteristic fusion network, and merging the two networks to obtain a defect detection network; finally, setting a loss function of the defect detection network, training hyper-parameters, and training the defect detection network to enable the baseline network, the multistage characteristic fusion network and a RPN (Risk Priority Number) to sharethe convolutional layer and calculated amount, thereby obtaining the completely trained defect detection network model and further detecting the steel plate surface defects. The steel plate surface defect detection method based on multistage characteristics of the convolutional neural network provided by the invention has strong defect classification ability, and specific types and accurate position information of the defects can be completely acquired. Moreover, configuration of hardware needed by detection is reduced.

The invention discloses a vehicle detection method based on a convolutional neural network. The method includes the step S1 of collecting vehicle samples and non-vehicle samples and classifying the vehicle samples, the step S2 of preprocessing the samples, the step S3 of training a CNN vehicle detector, the step S4 calculating an average similarity table of a characteristic pattern, the step S5 of constructing a similarity characteristic pattern set, the step S6 of obtaining a CNN-OP vehicle detector, the step S7 of obtaining detection images, the step S8 of preprocessing the obtained detection images, the step S9 of constructing an image pyramid for the detection images, the step S10 of extracting characteristics, the step S11 of scanning characteristic patterns, the step S12 of classifying the characteristics, and the step S13 of combining detection windows and conducting output. An offline optimization scheme is put forward, the convolutional neural network which is completely trained is optimized, the strategy of scanning the windows after extracting the characteristics is adopted at the detection stage, and therefore the characteristics are prevented from being repeatedly calculated, and the detection speed of the system is increased.

The invention discloses a clustering analysis and decision tree algorithm-based truck loading work time prediction model. A clustering analysis and decision tree mixed algorithm is introduced, factors influencing inventory control are abstracted out, related historical data serves as a training sample, and finally the truck loading work time can be effectively predicted by using a trained decision tree data model; and the historical data of truck loading is deeply mined by utilizing a data mining technology based on a demand, and an available, easy-to-use and high-accuracy data model is generated. The clustering analysis and the decision tree algorithm are combined and complement each other, so that the accuracy of the data model is improved; an optimization policy is adopted for an original decision tree algorithm under the condition of establishing a simple and accurate data model, so that the calculation amount is reduced and the algorithm efficiency is improved; and through the data model, a relatively accurate time interval of cargo loading can be predicted and used for better manual decision-making.

The invention relates to the field of indoor navigation and discloses an indoor location method and an indoor location system. In the invention, vision road signs (artificial or natural road signs) are deployed in fixed indoor positions. During the location process, a movable robot photographs images from surrounding environment, wherein the images are subjected to pretreatment and character extraction. A characteristic image after the character extraction is matched with the road signs in a road sign character library to obtain a road sign which can be used for location in the image. Through a vision distance measurement scheme on the basis of the characteristic road sign, real-time location is carried to the movable robot. In the embodiment, with the vision road signs (passive road sign), the road sign itself does not emit a wireless signal and has low cost. Through the vision distance measurement scheme, only one road sign is required to deploy within in vision range of the robot, so that the method and the system need fewer road signs, thereby reducing deployment cost of indoor navigation.

The invention claims an electroencephalogram recognizing method combing a convolutional neural network (CNN) with a long and short time memory network (LSTM). The method comprises the following steps:firstly, acquiring electroencephalogram signal data by using an Emotive acquiring instrument, and carrying out pretreatment such as mean removal, filtering and normalization on the acquired electroencephalogram signals; secondly, inputting pretreated data into a convolution layer and a pooling layer to extract space features; and finally, directly connecting the rear of the pooling layer to LSTM,extracting temporal order information of electroencephalogram data, and finishing a classifying task through Dropout and a fully connected layer. The temporal and spatial features of electroencephalogram signals can be fully utilized, the space and temporal order information of the electroencephalogram data are extracted, thus, the classifying accuracy of the electroencephalogram signals is improved, and a new way is provided for research on electroencephalogram recognition.

The invention provides a three-level midpoint potential balance control method based on zero sequence voltage injection, which is a control method capable of effectively controlling the balance of the midpoint potential in a high-power diode clamped three-level inverter. The control method comprises: collecting to obtain voltage values and three-phase output current values of two capacitors on the two sides of a direct current bus through voltage and current sampling circuits, calculating the voltage difference of two capacitors, judging whether the midpoint level is balanced and judging the relation of the three-phase output current and midpoint current during vector output under multi-carrier wave PWM modulation policy, calculating to generate a three-phase modulation voltage by using a zero sequence voltage injection method according to a zero sequence voltage selection principle, comparing with carrier waves to generate a needed switching sequence so as to achieve the purpose of controlling the midpoint level balance in the three-level inverter. The invention not only controls the midpoint level effectively, but also has the advantages of simple control method and strong robustness, and can ensure the inverter to work stably in a total range.

The invention discloses a depth extraction method of merging motion information and geometric information, which comprises the following steps: (1) carrying out scene segmentation on each frame of two-dimensional video image, and separating the static background and the dynamic foreground; (2) processing a scene segmentation chart by binaryzation and filtering; (3) generating a geometric depth chart of the static background based on the geometric information; (4) calculating the motion vector of the foreground object, and converting the motion vector into the motion amplitude; (5) linearly transforming the motion amplitude of the foreground object according to the position of the foreground object, and obtaining a motion depth chart; and (6) merging the motion depth chart and the geometricdepth chart, and filtering to obtain a final depth chart. The method only calculates the motion vector of the separated dynamic foreground object, thereby eliminating the mismatching points of the background and reducing the amount of calculation. Meanwhile, the motion amplitude of the foreground object is linearly transformed according to the position of the foreground object, the motion amplitude is merged into the background depth, thereby integrally improving the quality of the depth chart.

The invention relates to a driving assistance system and a real-time warning and prompting method for a vehicle. Image data of collected targets in front of a current vehicle is collected through an image collection module, then, a lane target in the image data is recognized by a traffic target recognition submodule, a region of interest is marked out according to the recognized lane target, and the region of interest is served as a key recognized target; and the image collection module is controlled by a data processing module to take a lane corresponding to the region of interest as an image collection target. Therefore, after the region of interest is mainly detected, the image collection module is capable of collecting image data required by a user more in time without collecting all forward data, therefore not only is collection capacity of the image collection module lowered but also data processing capacity of the data processing module is reduced, so that calculation amount of the whole system is reduced, calculation speed is accelerated, and warning response time is shortened.

The invention discloses a license plate detection method based on deep learning, comprising the following steps: using a fast-rcnn algorithm to train an RPN convolution neural network and a fast-rcnn convolution neural network; building an image library with marks and tags as a sample set; using the trained RPN convolution neural network to process the images in the sample set to get a rough license plate area; sending a rough license plate box to the trained fast-rcnn convolution neural network for judgment; and judging whether the rough license plate area is an optimal license plate area according to the output vector of the fast-rcnn convolution neural network, and if the rough license plate area is an optimal license plate area, taking the rough license plate area as a final license plate area. According to the invention, a multi-scale and multi-proportion reference box is adopted in training of the RPN convolution neural network, so detection of license plates with unconventional scale and proportion is promoted effectively. The RPN convolution neural network and the fast-rcnn convolution neural network share convolution layer parameters, so the whole system is simpler, less amount of calculation is needed, and the rate of missed detection is lower. Moreover, the real-time requirement of the system can be satisfied.

The invention discloses a gesture identification method and a gesture identification system, wherein the method comprises the following steps: step S1, a gesture detecting step: detecting a gesture in video stream in real time, marking a region in which the gesture is detected as an interested region; step S2, a gesture segmenting step: processing the interested region by utilizing skin color segmentation, and then performing edge detecting and outline extracting to obtain a point sequence of a hand-shaped outline; step S3, a gesture identifying step: firstly extracting a Fourier descriptor of the hand-shaped outline and then mapping the Fourier descriptor to a new vector of a characteristic space by utilizing PCA (principal components analysis), comparing the distance between the new vector and a gesture clustering center obtained by training, and judging a gesture type represented by the vector. The gesture identification method and the gesture identification system disclosed by the invention can improve the gesture identification accuracy and the gesture identification efficiency, and can effectively avoid background color interference.

The invention provides a joint measurement method based on a laser radar and a binocular visible light camera to obtain accurate and dense three-dimensional information simply and efficiently. The joint measurement method comprises assuming that the laser radar is a camera device with a fixed internal reference; directly projecting three-dimensional point cloud data into a two-dimensional image, calculating rotation and translation relationships between the laser radar device and the binocular camera by using an image processing method and the matching between the two-dimensional images; creatively introducing an idea for obtaining a matrix norm and a matrix trace to solve a rotation matrix; finally fusing laser radar and binocular stereo vision point cloud data. The method not only can obtain an accurate position and attitude information, but also can reconstruct the special texture and feature information of a target surface, which has a high application value for spacecraft dockingand hostile satellite capture in the military field and workpiece measurement and unmanned driving in the civilian field.

The invention provides a method and system for collecting aggregate information about network traffic, while maintaining processor load relatively constant despite substantial variation in network traffic, and capable of substantially accurate frequency measurement even for relatively infrequent events. A packet monitoring system includes an input port for receiving network packets, a sampling element for selecting a fraction of those packets for review, and a queue of selected packets. The packets in the queue are coupled to a packet-type detector for detecting packets of a selected type; the system applies a measurement technique for determining a frequency measure for those detected packets. The system includes a feedback technique for adaptively altering the sampling rate fraction, responsive to the queue length and possibly other factors, such as processor load or the detected frequency measure. The measurement technique also determines an error range and a measure of confidence that the actual frequency is within the error range of the measured frequency. The system can detect packets of multiple selected types essentially simultaneously, and provide measured frequencies and error ranges for all of the multiple selected types at once. Also, the measurement technique is selected so as to impose relatively light processor load per packet.

The invention discloses a clustering-based blind source separation method for synchronous orthogonal frequency hopping signals. The method comprises the following steps of: acquiring M sampled paths of discrete time-domain mixed signals; obtaining M time-frequency domain matrixes of the mixed signals; preprocessing the time-frequency domain matrixes of the frequency hopping mixed signals; estimating frequency hopping moments, normalized mixed matrix column vectors and frequency hopping frequency; estimating time-frequency domain frequency hopping source signals by utilizing the estimated normalized mixed matrix column vectors; splicing the time-frequency domain frequency hopping source signals between different frequency hopping points; and recovering time-domain source signals according to time-frequency domain estimate values of the source signals. According to the method, the frequency hopping source signals are estimated only according to the received mixed signals of a plurality of frequency hopping signals under the condition of unknown channel information, and the frequency hopping signals can be subjected to blind estimation under the condition that the number of receiving antennae is smaller than that of the source signals; short-time Fourier transform is utilized, so that the method is low in computation amount; and the frequency hopping signals are subjected to blind separation, and meanwhile, a part of parameters can also be estimated, so that the method is high in practicability.

The invention provides a prospective interpolation system for compressing and smoothening small segment paths. The prospective interpolation system comprises a path smoothening module, a curve scanning and segmenting module, a bidirectional acceleration module, a speed planning module and an interpolation module, wherein the path smoothening module is used for extracting small segment coordinate information in codes by reading information of G01 segments of numerical control codes G, conducting calculation according to the coordinate information and conducting path compression and smoothening on small segments; the curve scanning and segmenting module, the bidirectional acceleration module and the speed planning module complete S-shaped speed planning tasks with limited jerk; the interpolation module is used for generating discrete interpolation points according to planning speed obtained through calculation; the interpolation points can be stored and used for position closed-loop control. The prospective interpolation system is high in calculation efficiency in the whole process, small in calculation quantity, simple in programming realization and capable of being further applied to a high-speed and high-precision numerically-controlled machine tool.

The invention discloses an indoor scene 3D reconstruction method based on Kinect and solves the technical problem of the real-time reconstruction of an indoor scene 3D model and avoidance of excessive redundant points. The method comprises steps of: obtaining the depth data of an object by using Kinect and de-nosing and down-sampling the depth data; obtaining the point cloud data of a current frame and calculating the vector normal of each point in the frame; using a TSDF algorithm to establish a global data cube, and using a ray casting algorithm to calculate predicted point cloud data; calculating a point cloud registration matrix by using an ICP algorithm and the predicted point cloud data, fusing the obtained point cloud data of each frame into the global data cube, and fusing the point cloud data frame by frame until a good fusion effect is obtained; rendering the point cloud data with an isosurface extraction algorithm and constructing the 3D model of the object. The method improves the registration speed and the registration precision, is fast in fusion speed and few in redundancy points, and can be used for real-time reconstruction of the indoor scene.

The invention discloses an image processing method for automatic pointer-type instrument reading recognition. The method comprises the following steps: (1) Hough circle detection is carried out on the image, a weighted average method is used for positioning the circle center and the radius of a dial, and a dial region square image is extracted; (2) the image is pre-treated, and a binary thinning image of the instrument pointer is extracted; (3) a central projection method is used for determining a pointer angle; (4) a zero graduation line and full graduation line position templates are extracted, and a range starting point and ending point positions are calibrated; (5) by using template matching, zero graduation line and full graduation line angles are obtained; and (6) according to the pointer angle, the zero graduation line angle and the full graduation line angle, the pointer reading is obtained through calculation. Thus, the problem that the instrument dial position on the acquired image is not fixed as the relative position between a camera and the pointer-type instrument is not fixed can be solved, subjective errors as the reading of the instrument is read manually can be eliminated, the efficiency and the precision can be improved, safety of people is ensured, the application range is wide, and robustness is strong.

The invention discloses a voice wake-up method and device, wherein the voice wake-up method comprises the following steps: S1, acquiring a wake-up word, generating a junk word, which meets the preset condition, according to the wake-up word, and establishing a recognition network according to the wake-up word and the junk word; S2, acquiring voice information, which is inputted by a user, segmenting the voice information into a plurality of voice frames, and extracting acoustic features in each voice frame; S3, sequentially conducting likelihood computation on the acoustic features according to an acoustic model of a convolutional neural network to obtain the acoustic feature score of each acoustic feature; S4, selecting an optimal identification path from the recognition network according to the acoustic feature scores, and considering a voice result corresponding to the optimal identification path as an identification result; and S5, calculating the confidence corresponding to the identification result according to the identification result, and acquiring the wake-up result according to the confidence and outputting the wake-up result. The voice wake-up method and device, disclosed by the embodiment of the invention, have the advantages that the calculated amount is small, the power consumption is low, the false alarm rate of voice wake-up can be reduced, and the user's experiment is improved.

A method for simulating a physical system using finite element techniques, wherein two or more distinct models corresponding to distinct regions within the modeled system are solved, each with a corresponding evaluator. Nodes which lie on the boundaries between the models may have different values corresponding to the different models. When a particular model is solved, the evaluator for that model is used to obtain the appropriate values for each of these common nodes. In one embodiment, a first model is defined, then a region corresponding to a particular feature within the system is carved out of it. A finite element model corresponding to the feature is then inserted into the region. The finite elements may be adapted to share nodes on the boundaries between them.