171results about How to "Improve object detection accuracy" patented technology

The invention relates to a rapid target detection method based on a convolutional neural network, and relates to the computer vision technology. The rapid target detection method comprises the following steps: training convolutional neural network parameters by utilizing a training set; solving the problem of max-pooling losing feature by using an expander graph and generating a discriminative complete feature graph; regarding the full-connection weight of the convolutional neural network as a linear classifier, and estimating the generalization error of the linear classifier on the discriminative complete feature by using a probable approximately correct learning framework; estimating the required number of the linear classifiers according to the generalization error and the expected generalization error threshold value; and finally, completing the target detection on the discriminative complete feature graph by using the linear classifiers on the basis of a smooth window. The detection efficiency and the target detection precision are obviously improved.

A display includes: display pixel electrodes; common electrodes; a display layer; a display control circuit; touch detection electrodes; and a touch detection circuit detecting an external proximity object based on a detection signal obtained from the touch detection electrodes with use of a common drive voltage for display applied to the common electrode as a touch sensor drive signal. The touch detection circuit includes: a first filter allowing a fundamental detection signal, contained in the detection signal and having a frequency same as a fundamental frequency of the touch sensor drive signal, to pass therethrough, a plurality of second filters separately allowing two or more harmonic detection signals, contained in the detection signal and having frequencies same as respective harmonic frequencies of the touch sensor drive signal, to pass therethrough, and a detection section performing a detection operation based on the fundamental detection signal and the harmonic detection signals.

The invention requests to protect a multifunctional V2X intelligent roadside base station system. The system comprises roadside sensing equipment, an MEC server, a high-precision positioning service module, a multi-source intelligent roadside sensing information fusion module and a 5G/LTE-V communication module. An intelligent roadside device integrating C-V2X communication, environmental perception and target recognition, high-precision positioning and the like is designed, and the problem that multi-device information fusion and integration are inconvenient in intelligent transportation is solved. In the system, a C-V2X intelligent road side system architecture and a target layer multi-source information fusion method are designed. Road side multi-source environment cooperative sensing is combined, real-time traffic scheduling of the intersection is realized by using a traffic scheduling module in the MEC server, and communication and high-precision positioning services are providedfor vehicle driving, and finally the target information after fusion processing is broadcasted to other vehicles or pedestrians through a C-V2X RSU (LTE-V2X/5G V2X and the like) according to an application layer standard data format, so the driving and traffic safety is improved.

In order to solve the problems that in spaceborne optical remote sensing image ship target detection, the ship target scale change is large, and the calculated amount of a spaceborne platform is seriously limited, a lightweight multi-scale feature extraction network module is introduced, and the ship target detection efficiency of a deep learning network can be effectively improved. The inventiondiscloses a satellite-borne optical remote sensing image ship target detection method based on a lightweight receptive field pyramid. A method for constructing a lightweight receptive field pyramid byintroducing hole convolution is adopted, a multi-scale feature fusion detection module is constructed according to multi-scale features extracted from the receptive field pyramid, and the adaptability to optical remote sensing image ship target features is improved under the condition that the network scale is limited.

The invention discloses a vehicle detection method based on laser and vision fusion. The method comprises the following steps of 1) acquiring target detection information for an input image and laserpoint cloud; 2) performing optimal matching on the images of the front frame and the rear frame and the point cloud detection frame, and establishing a tracking sequence of an image and point cloud detection target; 3) fusing the tracking sequences of the image and the detection frame thereof and the tracking sequences of the point cloud and the detection frame thereof; 4) classifying all the target detection boxes, outputting a fusion list, and outputting a fusion result; and 5) obtaining the accurate position of the surrounding vehicle relative to the vehicle in the current frame, reading the next frame of image and the point cloud data, circulating the steps 1) to 5), and outputting a fusion detection result. According to the method, on the basis of point cloud and image target detection, the detection result is subjected to information tracking, the detection result is optimally matched, and the fusion result is preferentially input into the final fusion list, so that compared witha single sensor target detection method, the target detection precision is improved, and the false detection rate is reduced.

An object detection method and an object detection system, suitable for detecting moving object information of a video stream having a plurality of images, are provided. The method performs a moving object foreground detection on each of the images, so as to obtain a first foreground detection image comprising a plurality of moving objects. The method also performs a texture object foreground detection on each of the images, so as to obtain a second foreground detection image comprising a plurality of texture objects. The moving objects in the first foreground detection image and the texture objects in the second foreground detection image are selected and filtered, and then the remaining moving objects or texture objects after the filtering are output as real moving object information.

A radar apparatus is installed in a vehicle that moves along its direction of travel. A radar transmission unit transmits a high frequency radar transmission signal from a transmit antenna in each transmit period. In a radar reception unit, antenna system processing units each generate a correlation vector by computing the correlation between reflected wave signal from a stationary object or a moving object and the radar transmission signal. A Doppler frequency-azimuth conversion unit converts Doppler frequencies into the components of an azimuth in which the stationary object is present using an estimated vehicle speed vector for the vehicle. A stationary object azimuth estimation unit generates the power profile of the reflected wave signal using the correlation vector and a direction vector corresponding to the components of the azimuth in which the stationary object is present.

The invention discloses a regional average value kernel density estimation-based moving target detecting method in a dynamic scene. The method comprises the following steps of: firstly, initializing a background model; secondly, building a time and space background model for describing the dynamic complex scene by using a training sample in a background modelling process and considering the time sequence characteristics of pixel points in a video frame and the space characteristics in the adjacent regions of the pixel points; thirdly, continuously updating the background model by using the new video frame sample in a moving target detecting process; fourthly, adapting to the instantaneous background change by the regional kernel density estimating method and adapting to the continuous background change by using single Gauss background model, wherein the combination of the two models can fast and accurately adapt to the continuous change of the background and increases the executing efficiency of the method at the same time; and finally performing a foreground detecting method by providing an adjacent region information amount-based method so as to further remove noise points and inanition of a moving target in the background region in the detecting process and more completely extract the moving object in the foreground. The method can be widely applied to alarming the suspicious moving target in an intelligent monitoring system in an outdoor scene or a prohibited military zone and has wide market prospect and application value.

Disclosed is an object detection device capable of improving object detection accuracy by preventing erroneous combination detection. In this device, a detection object region correction unit (103) corrects a detection object region set by a detection object region set unit (102) on the basis of moving speed map information associated with a coordinate group in a reference image plane and the detected moving speed on each coordinate, which is detected by a radar. Thus, the detection object region can be amended by the use of the moving speed map information even when the detection object region is obtained as a result of erroneous combination. As a result, the object detection accuracy can be improved.

The invention discloses a salient visual method based on polarization imaging and applicable to underwater target detection, comprising the following steps: (A) acquiring auto-registration polarization images of the same underwater position at multiple angles; (B) performing underwater image restoration based on polarization information; (C) extracting global texture features; (D) extracting color features based on global contrast; (E) fusing visual saliency features; (F) performing saliency map optimization and target extraction based on the target center and the gray gravity center; and (G) performing threshold segmentation on a final saliency map to realize underwater target detection. Salience optimization is realized by use of the target center probability, image gray gravity center and space smoothness. The background is further restrained, and the foreground is highlighted. High detection rate and high identification rate are achieved for target detection in a complicated water environment, and the real-time requirement is satisfied. The method has a good application prospect.

Exemplary computer-implemented method and system can be provided for constructing a model in a neural network for object detection in an unprocessed image, where the construction can be performed based on at least one image training batch. The exemplary model can be constructed by training one or more collective model variables in the neural network to classify the individual annotated objects as a member of an object class. The exemplary model, e.g., in combination with the set of specifications when implemented in a neural network, can perform object detection in an unprocessed image with probability of the object detection.

The invention discloses an improved target detection method based on a residual network, based on YOLO V3-tiny network, features are extracted through continuous convolution operation, finally, pictures are divided into 13*13 grids, for each grid unit, a detection frame of a target with a center point falling into the grid unit is predicted through three anchor frames. The method specifically comprises the following steps of determining the number of types of the target to be identified to form a data set; establishing a target detection neural network; obtaining training weight files. Throughthe lightweight target detection network YOLOV3-tiny, the calculation amount is small, the target detection task can be carried out in the embedded hardware, and the real-time performance is ensured.According to the method, the original feature extraction network is replaced by the residual network resnet18, and for the feature extraction network with the same layer number, the feature extraction capability of the network can be improved by adding the residual structure by the residual network, so that the target detection precision can be improved on the premise that the detection speed isnot reduced.

A display device with high accuracy in object detection is provided. The display device includes a light-detection touch sensor, a capacitive touch sensor, and an illuminance sensor configured to detect the illuminance of external light. The information about the illuminance detected by the illuminance sensor is used to choose either the light-detection touch sensor or the capacitive touch sensor for imaging. That is, an appropriate touch sensor is chosen from the two kinds of touch sensors. Accordingly, the object detection accuracy can be prevented from decreasing due to the influence of external light.

The invention discloses a YOLO target detection method using GPU hardware acceleration. The method comprises steps of: (1) initializing a convolutional neural network (CNN); (2) acquiring a training sample; (3) determining the grids of the training sample; (4) training the CNN; (5) determining whether a loss value is less than 0.01, saving the trained CNN if so, or acquiring and training a next training sample if not; (6) saving the model of the trained CNN in a computer hard disk; (7) extracting the characteristics of a test picture; (8) determining the location rectangular frame of the testpicture target; and (9) ending the target detection. The method can realize feature extraction on a target in the image on a general computer, then marks the location of the target with the location rectangular frame, and marks the category of the target at the upper right corner of the location rectangular frame.

In a system for detecting a target object, a similarity determining unit sets a block in a picked-up image, and compares a part of the picked-up image contained in the block with a pattern image data while changes a location of the block in the picked-up image to determine a similarity of each part of the picked-up image contained in a corresponding one of the different-located blocks with respect to the pattern image data. A specifying unit extracts some different-located blocks from all of the different-located blocks. The determined similarity of the part of the picked-up image contained in each of some different-located blocks is equal to or greater than a predetermined threshold similarity. The specifying unit specifies, in the picked-up image, a target area based on a frequency distribution of some different-located blocks therein.

The invention discloses a tunnel pedestrian target detection method based on a cascaded super-resolution network and an improved Faster R-CNN, and the method comprises the following steps of S1, training the super-resolution network, and obtaining an SRCNN super-resolution network model; S2, acquiring a tunnel pedestrian training sample and marking the pedestrians; S3, clustering the dimension proportion of a label box, and selecting a proper anchor frame dimension in the RPN network; S4, training the Faster R-CNN network, and obtaining a trained model; and S5, detecting the tunnel pedestriantarget by adopting the trained model to obtain a detection result. Compared with an original Faster R-CNN network, the method has the higher detection precision and can be effectively applied to low-resolution pedestrian target detection in a tunnel environment.

There is provided an image processing system in which an image capture apparatus and an image processing apparatus are connected to each other via a network. When a likelihood indicating the probability that a detection target object detected from a captured image is a predetermined type of object does not meet a designated criterion, the image capture apparatus generates tentative object information for the detection target object, and transmits it to the image processing apparatus. The image processing apparatus detects, from detection targets designated by the tentative object information, a detection target as the predetermined type of object.

The invention discloses a robot disordered grabbing method and system based on machine vision. The method comprises the steps that a kinect camera and a grabbing system of a robot are built, the kinect camera is used for collecting image data of the surface of a target object, and the image data are preprocessed to obtain three-dimensional point cloud data; target detection, target segmentation, target clustering, key point feature extraction, feature registration and target object recognition are performed on the acquired three-dimensional point cloud data to obtain pose information of the target object; and according to the calibration result of the hand-eye system, coordinate conversion is carried out on the pose information, and a control instruction is sent to a robot to grab the target object. By analyzing the defects of the algorithm in the disordered grabbing process of the robot and improving the existing algorithm, the complex environment where the robot is located is adapted, and therefore the disordered grabbing accuracy of the robot and the flexibility of the robot are improved.

The invention discloses a target detection method, system, and device and a storage medium based on an area proposal. The method comprises the steps of inputting an image to be detected into a targetdetection network, receiving a final boundary frame outputted from the target detection network, and determining a target to be detected from the image to be detected according to the final boundary frame. The invention provides a novel target detection network, A target detection network includes a plurality of branches, Each branch contains local information and global information, and each branch continues to extract and learn feature information based on the processing results of the previous branch, so it can give attention to both local information and global information of the image, and can achieve high accuracy of target detection. The invention is widely applied to the technical field of image recognition.

Disclosed is an apparatus and method for providing information regarding a blind spot in a vehicle. The apparatus includes a view transforming area detector that is configured to detect a predefined side area and rear side area from a captured image input from a side imaging device. The imaging device is configured to capture the image including the blind spot of the vehicle. Additionally, the apparatus includes a view transformer that is configured to view transform an image of the side area and an image of the rear side area based on a pre-set view transformation parameter and generate view transformed images corresponding to the images of the side area and the rear side area.

The present disclosure discloses an object detection method based on machine learning. The object detection method senses a motion of an object by a motion sensor to generate a testing image, captures the testing image by an image sensor to transmit a sensed image to an object detection device, and detects a position of the object in the sensed image by the object detection device. Therefore, the present disclosure increases the accuracy of image recognition under various conditions.

The invention relates to a target detection method based on global and local information fusion, and belongs to the field of video image processing. The method comprises the following steps: firstly,sending a scene into a convolutional neural network to increase the memory ability of the network, so that the network better learns scene context information to obtain global scene features; secondly, establishing a relationship between objects in a self-adaptive manner by referring to an attention mechanism to obtain local object characteristics; and finally, fusing scene features and object features through information transmission to enhance feature expression. The method has the advantages that global scene features and local object features are considered at the same time, target features are better represented through information transmission, and a large number of contrast experiments show that the detection performance of the method is obviously superior to that of other target detection methods.

The invention relates to a three-dimensional target detection system based on laser point cloud; the system comprises a voxel size division module, a feature coding module, a feature extraction and fusion module, a target regression and detection module and a laser radar. The output end of the laser radar is connected with the input end of the target regression and detection module through the voxel size division module, the feature coding module and the feature extraction and fusion module in sequence, and during use, firstly, the voxel size division module performs voxel division on a three-dimensional target point cloud obtained from the laser radar by adopting different voxel scales; a plurality of voxelized point clouds are obtained, then feature coding is performed on the plurality of voxelized point clouds by a feature coding module; feature extraction and fusion are performed on the coded voxelized point clouds by a feature extraction and fusion module to obtain a final feature map; and finally, a three-dimensional target detection box is obtained by a target regression and detection module according to the final feature map. The design can guarantee that the structural features of the point cloud are not lost, and improves the detection precision of the three-dimensional target.

The invention discloses a picking point positioning method which comprises the following steps: obtaining a trained YOLOv3 target detection model, collecting and detecting a fruit growth area image, obtaining the number and positions of fruits in a detection view field, and fusing a feature extraction network of dense connection and residual idea with the target detection model; judging the numberof fruits in the visual field, and judging the images as a distant view and a close view according to the number; performing branch segmentation on the close-range image by using a semantic segmentation model to obtain a branch segmentation image; carrying out target detection on the distant view; judging whether the number of fruits in the branch segmentation image is 1 or not, if yes, using a single-fruit picking strategy, and if not, using a multi-fruit picking strategy; and acquiring final picking positioning point. According to the method, non-damage picking of mature bunches of fruits can be achieved, the shear point positions of the mature bunches of fruits are accurately found and positioned through an algorithm, fruit shear type picking is achieved, and integrity and picking efficiency of the fruits are guaranteed.