68results about How to "Implement object detection" patented technology

The invention provides a target detection method and device and a fuzzy processing method and device. The target detection method comprises the following steps: the multimedia information which is collected by the main acquisition equipment of at least two sets of acquisition equipment and includes at least two target objects is acquired; If it is detected that the multimedia information collectedby the main acquisition equipment does not meet the preset detection condition, the multimedia information collected by each set of acquisition equipment using different acquisition parameters is acquired; and at least two target objects are detected according to the multimedia information collected by all the acquisition equipment. The detection accuracy of the target objects can be enhanced.

The invention discloses a multichannel radar interference inhibition and then target detection method during coexistence of clutter and interference. Interference inhibition, clutter inhibition and CFAR (constant false alarm rate) detection area realized on the basis of the idea of multichannel adaptive detection. The number and direction of interference are obtained by means of reconnaissance pulses in a rest period of a radar, singular value decomposition is carried out on an interference guiding matrix, and an interference inhibition matrix is constructed; the interference inhibition matrixis used to carry out interference inhibition on to-be-detected data and a training sample, the dimension of data is reduced, and requirement for the number of training samples during subsequent adaptive detection is lowered; according to a generalized likelihood ratio criterion, the to-be-detected data and training sample after interference inhibition are combined to detect design of a detector;and a detection threshold is determined according to statistical characteristic of the detector and the false alarm rate set by the system, and compared with a detection statistic quantity of the detector, if the detection statistic quantity is greater than the threshold, it is determined that there is a target, and otherwise, it is determined that there is no target. Thus, interference inhibition, clutter inhibition and CFAR detection can be realized at the same, time, and work is normal when the number of training samples is lower than the number of system channels.

The invention discloses an underwater image object detection method, and the method comprises the steps: respectively extracting the color, gray scale and depth information of an underwater image as the input of an underwater image target detection model; respectively calculating the global contrast of the color, gray scale and depth information of each pixel of the underground image in the underwater image target detection model, carrying out the fusion, and generating multi-information fusion global contrast; determining that a point belongs to an image region where the a target is located when the multi-information fusion global contrast of the pixel point is greater than a certain threshold value, so as to discriminate the image region where the target is located, thereby achieving the detection of an underground image target. The method enables the image depth information to the underground image target detection, can solve a bottleneck problem that a target is difficult to detect in underwater high-scattering and strong attenuation environments through the fusion of color and gray scale information, and accurately detects the underwater image target.

The invention discloses a target detection method and device, and the method comprises the steps: obtaining an initial image, an initial boundary frame being marked for an interest target in the initial image; inputting the initial image into a pre-trained regression model to obtain a to-be-segmented image which is output by the regression model and comprises an adjusted boundary frame, the regression model being obtained by pre-training a training image marked with a real boundary frame of the interest target and marked with the initial boundary frame of the interest target; and segmenting the image contained in the adjusted boundary frame in the to-be-segmented image as a to-be-classified image containing the interest target. According to the method and the device, the marking precisionof the boundary frame of the interest target is improved.

The invention discloses an image background removal method fusing color and local ternary similar mode characteristics. The method comprises steps that S1, a designated target image is acquired, color characteristics of each pixel of the acquired image are respectively calculated, after ternary division of each pixel and an adjacent pixel of the present image is carried out based on LTSP characteristics, intra-LTSP characteristics corresponding to each pixel are acquired through calculation, background modeling is carried out for the target image based on the color characteristics and the intra-LTSP characteristics, and a background model is acquired through establishment; and S2, the established background model is utilized to carry out foreground segmentation of the target image to remove the background image, and model update of the background model is carried out according to the color characteristics of each pixel and the intra-LTSP characteristics. The method is advantaged in that the method is simple for realization, cost is low, adaptability to illumination change of the background is strong, background removal efficiency is high, the effect is good, and the detection missing rate and the false detection rate are low.

The invention relates to a system and method for establishing target division remote damage assessment of different vehicle types based on an artificial intelligence semi-supervised learning BIRCH method and belongs to the vehicle damage assessment field. The objective of the invention is to solve problems in target detection of collision vehicles after a collision. According to the technical schemes of the invention, a target detection subsystem is adopted to judge collision objects in the vehicle collision; and the target type detection subsystem learns target training data so as to generate a target model, wherein the target model is built by adopting the intelligent semi-supervised learning BIRCH algorithm. With the system and method provided by the technical schemes of the invention adopted, target detection in the vehicle collision can be realized; and a machine learning method is used in the remote damage assessment technical field, so that the accuracy of judgment in a damage assessment process can be improved.

The invention discloses a radar sequence signal detection method and system based on LSTM. The method comprises the following steps: acquiring radar signal data of a to-be-detected area; calculating the amplitude of the radar signal data; dividing the radar signal data into a training data set and a test data set; constructing an LSTM model; training the LSTM model through the training data set toobtain a trained LSTM model; predicting the test set through the trained LSTM model to obtain a predicted value; calculating a relative error between the predicted value and an actual value, whereinthe actual value is the amplitude of the radar signal data; and detecting the radar sequence signal according to the relative error. According to the method, subsequent moment signals are predicted through the trained LSTM model, points with large relative errors are regarded as abnormal values by comparing the relative errors of predicted values and actual values, radar target detection can be achieved, and clutter influences can be reduced.

The invention provides a target detection system based on a micro aircraft. Video data is acquired through a camera arranged on a four-axis flying platform; and then, a Raspberry pi 4 micro-control unit and a Call USB acceleration rod additionally arranged on the Raspberry pi 4 micro-control unit are used for bearing a target detection system comprising a multi-branch depth separable convolutionalneural network and a Single Shot MultiBox Detector operation module to perform object detection on a frame picture in the video data. According to the invention, the size of the model is reduced by using depth separable convolution, and the generalization of the model is improved by using a multi-branch structure. According to the invention, under the condition that a Coral USB acceleration rod is additionally arranged on Raspberry pi 4, an object can be rapidly detected through a constructed MBDSCNN-SSD target detection model.

The invention discloses a zebra crossing detection method and system for assisting safe driving of a vehicle. The method includes: S1, shooting a target scene by a vehicle-mounted camera, and acquiring a road image in the target scene; S2, performing line feature detection on the road image to obtain Line feature; S3, obtain the position information of the vanishing point according to the line feature; S4, carry out secondary line feature detection to the road image, obtain the straight line feature in the road image; S5, black and white the line feature in the road image mark, and obtain all line segment information; S6, perform a screening on the line segments in the road image to obtain candidate ROIs; S7, perform secondary screening on candidate ROIs, and obtain the final detection result. The beneficial effect of the invention is that not only a single target can be detected, but also multiple targets to be detected under complex intersections can be detected.

The present invention provides a target detection method based on a micro air vehicle. The present invention reduces the size of the model by using depth-separable convolution, and improves the generalization of the model by using a multi-branch structure. The target detection model of the present invention is further implemented in the front-end network MBDSCNN Additional convolutional layers are added to the structure, and the size of the additional convolutional layers is set to decrease layer by layer, so that the model can predict under multi-scale conditions. The target detection model established by the present invention can combine the output of each extra convolutional layer to replace the fully connected layer of MBDSCNN, and add a non-maximum suppression layer (NMS) after the fully connected layer to output the detection result. Therefore, compared with the prior art, the present invention can accurately identify targets at various scales, and accurately output detection results.