81results about How to "Reduce the number of parameters" patented technology

The invention discloses a multi-scale image semantic segmentation method. The method comprises the following steps: obtaining a to-be-segmented image and a corresponding label; constructing a full convolutional deep neural network, wherein the full convolutional deep neural network comprises a convolution module, a hole convolution module, a pyramid pooling module, a 1 * 1 * depth convolution layer and a deconvolution structure; setting hole convolution as channel-by-channel operation, and utilizing low-scale, medium-scale and high-scale characteristics in a targeted mode; training the full convolutional deep neural network, establishing a loss function, and determining parameters of the full convolutional deep neural network by training the sample image; and inputting the to-be-segmentedimage into the trained full convolutional deep neural network to obtain a semantic segmentation result. By means of the method, the image semantic segmentation problem with complex details, holes andlarge targets can be well solved while the calculated amount and the parameter number are reduced, and the consistency of category labels can be reserved while the target edges can be well segmented.

The invention discloses a remote sensing image ground object classification method based on superpixel coding and convolution neural network, using adaptive superpixel coding and double channel convolution neural network. The remote sensing image ground object classification method based on superpixel coding and convolution neural network includes the steps: utilizing a superpixel algorithm to perform image pre-segmentation; using a cluster method to merge neighboring and similar superpixel blocks, setting the size of the taken blocks, constructing three double channel convolution neural networks with different input size; inputting samples with different taken block size into the corresponding network; using the convolution neural networks to extract the data characteristics of two sensors respectively; merging the extracted characteristics for classification; and according to the size of the merged pixel block, determining the size of the taken blocks of the samples, and realizing adaptive selection of the utilized neighborhood information. The remote sensing image ground object classification method based on superpixel coding and convolution neural network can realize adaptive selection of the utilized neighborhood information to enable the neighborhood information to realize positive feedback effect and preferably utilize the neighborhood information to send the samples to different networks according to the neighborhood information so as to enable the samples with similar distribution to enter the same network, thus effectively improving the classification accuracy.

The invention provides a GNSS multimode single-frequency RTK cycle slip detection method and apparatus. The method comprises the following steps: communicating with a GPS satellite, a GLONASS satellite, a Galileo satellite and a Beidou satellite and obtaining corresponding data; according to a formula, carrying out calculation so as to obtain a residual error vector of a station epoch secondary difference carrier wave observation value of each satellite; according to the residual error vector, calculating an RMS value, and if the RMS value is greater than or equal to a threshold EPS, determining that a satellite generates a cycle slip; according to a formula (2), calculating a standard residual error V<->; comparing |error V<->| with u<alpha / 2>, and carrying out corresponding operation; and fusing the determined cycle slip with a cycle slip detected by use of a Doppler integration method. The invention brings forward a novel method for detecting a cycle slip by combining a multimode station epoch secondary difference method and a Doppler integration method. The multimode station epoch secondary difference method detecting the cycle slip based on a residual error domain only sets one station epoch relative clock error parameter, such that even if a single system has only one satellite, effective utilization and cycle slip detection can still be realized.

The invention relates to a KNN (K-Nearest Neighbor) sorting algorithm based method for correcting and segmenting the grayscale nonuniformity of an MR (Magnetic Resonance) image, belonging to the field of image processing. The method comprises the following steps of: firstly constructing a grayscale nonuniform field model by utilizing surface fitting knowledge and using a group of orthonormalization basis functions, and establishing energy functions; and then solving model parameters according to an energy function minimization principle to realize grayscale nonuniformity correction and image segmentation, wherein subordinate functions are solved by adopting an iterative algorithm and the KNN algorithm in the model parameter solving process, therefore a partial volume effect is greatly reduced while a grayscale nonuniform field is eliminated, and the influence of noises on the correction and the segmentation of the grayscale nonuniformity of the MR image is reduced. The subordinate functions are solved with KNN through the following steps of: firstly acquiring an accurate smooth normalization histogram by using a kernel estimation algorithm; then respectively solving a threshold value TCG between cerebrospinal fluids and gray matters and a threshold value TGW between the gray matters and white matters by using a maximum between-cluster variance method; carrying out rough sorting on the KNN sorting algorithm by utilizing the two threshold values; and finally accurately sorting points to be fixed by adopting the traditional KNN sorting algorithm.

The invention discloses a semantic segmentation method based on a pyramid cavity convolution network, and the method comprises the following steps: obtaining a medical image data set containing a realsegmentation result, and carrying out the preprocessing operation of data enhancement and the like of the data set; processing the preprocessed image through a residual recursion convolution module and a pooling layer to obtain shallow image features; obtaining deep image features through a network in which a pyramid pooling module and a cavity convolution module are connected in parallel; decoding the features of the deep image through a deconvolution layer, jump connection and residual recursion convolution module; inputting a decoding result into a softmax layer to obtain a category to which each pixel belongs; training a pyramid cavity convolution network, establishing a loss function, and determining network parameters through training samples; and inputting a test image into the trained pyramid cavity convolutional network to obtain a semantic segmentation result of the image. According to the method, multi-scale semantic information and detail information can be effectively extracted by adopting a hole convolution and pyramid pooling method, and the segmentation effect of the network is improved.

The invention relates to a test apparatus and method for uncorrelated sound vibration load combination application and uncorrelated multisource frequency domain load identification, and three methods for performing uncorrelated multisource frequency domain load identification in a complicated sound vibration simulation environment by utilizing the apparatus. The three methods are a general-reversion-of-least-square method, an improved regularization method and a multi-input multi-output support vector regression method respectively. The three methods are all capable of identifying the values of a plurality of uncorrelated frequency domain load sources at the same time according to vibration responses of a plurality of test points. The general-reversion-of-least-square method does not need to measure a phase of a transfer function; each frequency corresponding to the improved regularization method has an optimal regularization parameter; and the multi-input multi-output support vector regression method is capable of effectively avoiding the over-learning phenomenon.

The invention is applicable to the field of vehicle-mounted all-round view, and provides a vehicle-mounted three-dimensional all-round display method, a computer readable storage medium and a system.The method comprises the following steps: establishing a three-dimensional grid model, and amplifying the three-dimensional grid model to world coordinates; acquiring fisheye images around the vehiclebody collected by a camera, and performing distortion correction to obtain corrected images and coordinates of an imaging center; calculating to obtain a camera internal reference matrix according tothe coordinates of the imaging center; determining an external parameter matrix of each camera under a vehicle body coordinate system; determining a mapping relation between each point in the three-dimensional space and a planar pixel coordinate; rendering the content of the corrected image into a three-dimensional grid model to realize a three-dimensional panoramic all-round view; and displayingthe three-dimensional panoramic all-round view. The stereoscopic environment around the vehicle can be displayed on the display window more clearly, the view field range of the vehicle-mounted stereoscopic panoramic display system is widened, and the safety performance is improved.

The invention discloses an improved AlexNet-based burning arc identification method. The method comprises the following steps of S1, building a convolutional neural network and obtaining a training model; S2, obtaining an image; and S3, inputting the obtained image to the training model for performing image identification, wherein the convolutional neural network is improved AlexNet, the structureof the improved AlexNet sequentially comprises an input layer, multiple convolutional layers connected in sequence, a full connection layer and an output layer, the convolutional layers connected with the input layer adopt MXM convolutional kernel architectures, and the test of the convolutional layers adopt 1XM and MX1 convolutional kernel architectures. According to the method, MXM convolutional kernels in original AlexNet are replaced with 1XM and MX1 convolutional kernels connected in series, so that a parameter quantity is greatly reduced; a nonlinear layer is added, so that the networkstructure is deeper; and the model is simplified, so that the training time is greatly shortened, the training efficiency is improved, and the identification accuracy is higher than that of the original AlexNet.

The invention discloses a real-time point cloud model classification method based on a lightweight network LightPointNet, comprising the following steps: S1, analyzing characteristics of a deep convolution neural network structure, designing a lightweight real-time point cloud network LightPointNet according to application requirements, wherein the lightweight real-time point cloud network LightPointNet at least contains an input layer, convolution layers, a full connection layer and an output layer, the last convolution layer contains one pooling layer, and maximum pooling is adopted in eachcharacteristic channel carrying out convolution in the last convolution layer for generating a characteristic value; S2, inputting data, convoluting the data by virtue of three convolution layers, andobtaining the final characteristic values; S3, inputting the obtained characteristic values into the full connection layer to be classified; and S4, obtaining the final classification accuracy rate by utilizing a classification operation in the step S3. The method disclosed by the invention has the advantages of fewer network layers, fewer parameters, high processing speed and excellent classification performance.

The invention discloses a method for utilizing PMU (power management unit) disturbance data to carry out on-line identification on a synchronous generator q shaft parameter. The derivation process of an identification theory is as follows: an error objective function of q shaft parameter identification is obtained according to a generator five-stage model, then, the relative sensitivity of the objective function for each parameter of a q shaft is obtained by aiming at the unit of a practical electric network, a unit is divided into two types, namely an A type and a B type according to the characteristics of the parametric sensitivity of different units, and a method used for identifying the q shaft parameter of the A type unit is provided. The identification method provided by the invention is simple and is less in time consumption, and can be used for improving the identification precision of synchronous reactance.

The invention discloses an aerial photography car detection method based on YOLOv4, and the method comprises the steps: employing a YOLOv4 model of a darknet version, carrying out the labeling of cars with 2000 images in an open source aerial photography data set, arranging the cars into a data set format needed by the YOLOv4, and then employing the YOLOv4 to pre-train a weight to train an own aerial photography data set; L1 loss is used to reduce neural network weight to perform sparse training, a channel pruning technology is used to perform pruning training on YOLOv4, the model is compressed, the training speed is improved, the pruned weight is finely adjusted, and a random multi-scale training technology is used to improve the generalization ability of the training model, so that the precision is improved. And finally, the aerial image test set is tested by using the self-trained weight, so that the memory occupation space is reduced on the premise that the detection speed meets the real-time requirement and the detection precision is not influenced.

The invention provides a multiple isomerous water environment monitoring data evaluating and early-warning method, and belongs to the field of water quality information processing. The method of background deduction is used in extraction of a moving target, a frame number counter is installed at each pixel, and when one pixel is judged to be a foreground point, the frame number counter at the one pixel adds 1; when one pixel is greater than a set threshold value, the gray value of the one pixel servers as the gray value of the one foreground point; the percentage of the number of the foreground point pixels in all the pixels is a characteristic parameter of the moving target. In water-surface main characteristic parameter extraction, an image is divided into 16*16 macro blocks, the macro blocks with the percentage of the gray level having the maximum number of the pixels in all the pixels exceeding 40% are chosen, statistics is conducted on the area having the maximum number of alternative macro blocks, and the average gray value of all the macro blocks in the area is used as a water-surface main area characteristic parameter. Then, the two characteristic parameters and water quality attribute data collected by sensors are used as characteristic vectors, and the relation between the characteristic vectors and a water environment safety level is built. The multiple isomerous water environment monitoring data evaluating and early-warning method can detect the moving object when a background changes severely.

The invention discloses a transformer state evaluation method and system based on a principal component analysis method and a support vector machine. Firstly, a principal component analysis (PCA) method is adopted to carry out dimension reduction processing on state information with complex quantity of transformers, relevance among data is eliminated, most valuable key parameters are selected, andtherefore complexity of subsequent calculation is reduced. Subsequently, the LS-SVM serves as a classifier, the key parameters serve as input, parameter optimization is conducted through the particleswarm algorithm, and the state level of the transformer is evaluated. Meanwhile, a multi-level SVM system is provided, the one-to-many classification problem of the SVM is solved, and a state evaluation model of the transformer is established. According to the method, the parameter number is reduced, the calculation complexity is reduced, meanwhile, the state judgment accuracy of the transformeris guaranteed, and a good guarantee is provided for safe operation of a power system.

The invention relates to a second-order hybrid construction method and system for a complex-valued forward neural network. The second-order hybrid construction method comprises the following steps: initializing the structure and parameters of the complex-valued neural network according to a given task; adjusting parameters in the complex-valued neural network by using a complex-valued second-orderhybrid optimization algorithm, and judging whether a construction termination condition is met or not; verifying the generalization performance of the complex-valued neural network, storing the number of current hidden layer neurons and all parameter values of the complex-valued neural network, judging whether the adding standard of the hidden layer neurons is met or not, if yes, adding one hidden layer neuron to the current model by utilizing a complex-valued increment construction mechanism, and otherwise, adding one hidden layer neuron to the current model by utilizing a complex-valued increment construction mechanism; calculating a new hidden layer output matrix and an error function on the basis of current training, and returning to the previous step; if not, directly returning to the previous step; and further finely adjusting the parameters by using the complex-valued second-order hybrid optimization algorithm to obtain an optimal complex-valued neural network model. Accordingto the invention, the complex-valued neural network model with a reasonable structure can be constructed automatically.

The invention discloses a levelling and calibrating method used for an indoor space surveying and positioning system. The levelling and calibrating method comprises the following steps: rigidly connecting an inclinometer with a base; fixing a launch station at the top end of the base; fixing a laser at the top end of the launch station; supplying power by the launch station to drive the laser at the top end to emit visible laser to the outside, thus forming a scanning plane and guaranteeing that all points on the scanning plane are positioned inside a view field of a theodolite; reading the coordinate value in the Z-axis direction by utilizing three points which are aligned by the theodolite and are separated by 120 degrees; calculating the differences of three average values until any two difference value is smaller than 0.1mm, otherwise, adjusting the height of a support leg corresponding to the triangular base; and setting the output angle values of the two axes of the theodolite to be zero, and taking the output angle values as the levelling criterion of the launch station. According to the levelling and calibrating method, the calibration process of external parameters is simplified, and the applicability of the system is improved; a rotating shaft of the launch station is taken as the criterion of a geodetic coordinate system, and further, the position and posture relationship with other measuring instruments is established, so that the combined use of various instruments is convenient.

The invention discloses a double-frequency power amplifier digital predistortion device and method based on a piecewise linear function, which comprises a first digital pre-distorter, a second digitalpre-distorter, a first digital-to-analog converter, a second digital-to-analog converter, a first low-pass filter, a second low-pass filter, a first modulator, a second modulator, a first power divider, a power amplifier, an attenuation coupler, a second power divider, a first band-pass filter, a second band-pass filter, a first demodulator, a second demodulator, a third band-pass filter, a fourth band-pass filter, a first analog-to-digital converter, a second analog-to-digital converter and a predistortion training module. The invention also discloses a double-frequency power amplifier digital predistortion method based on the piecewise linear function. The device and the method in the invention reduce the number of parameters in the existing 2D-DPD double-frequency power amplifier predistortion model, thereby reducing the implementation difficulty and consumed computing resource of a model in the FPGA and accelerating the computing time of the optimal model parameter.

The invention provides a three-dimensional particle category detection method and system based on a convolutional neural network. The method comprises the following steps: constructing a three-dimensional mixed-scale dense convolutional neural network comprising a mixed-scale three-dimensional extended convolutional layer, dense connection and a loss function, training the convolutional neural network by using a three-dimensional frozen electron tomography image marked with the particle coordinates to obtain a particle selection model, and training the convolutional neural network by using thethree-dimensional frozen electron tomography image marked with the particle category to obtain a particle classification model; acquiring the three-dimensional frozen electron tomography image through a sliding window to obtain to-be-detected three-dimensional reconstructed subareas, predicting each subarea through the particle selection model, and combining prediction results of the subareas toobtain coordinates of each particle in the three-dimensional frozen electron tomography image; and extracting a three-dimensional image of each particle according to the coordinate of each particle, and inputting the three-dimensional image of each particle into the particle classification model to obtain the category of each particle.

The invention discloses a flower recognition method based on a convolutional neural network (CNN) with a ReLU activation function, which belongs to the technical field of image recognition. The method comprises steps: basic parameters of the CNN are set; the weight and the bias term are initialized, and convolutional down sampling layers are designed layer by layer; a random sequence is generated, 50 samples are selected at each time for batch training, a forward process, error conduction and a gradient calculation process are completed, and a gradient sum is updated to a weight model for next weight updating; and a well-set training function and an updating function are called for training, and the accuracy of a sample is tested. Flower recognition can be carries out effectively at a high speed under influences of lighting, rotation and occlusion conditions.

The invention relates to a high-order interactive multi-model filtering method based on mixture transition distribution. In order to solve the problems that in an existing method, high-order Markov chain setting parameters are multiple, setting processes are complex and the precision is low, the method comprises the steps that 1, a mixture transition distribution model is adopted for obtaining n-order model sequence transition probability; 2, real-time processing is performed at the moment k; 3, initialization is performed in the state when k is equal to 1; 4, initialization is performed in the state when k is equal to 2; 5, judgment is performed on k, when k is equal to n, initialization is performed on the n-order model sequence probability at the moment k, an estimated value of the n-order model sequence state and corresponding covariance when k is equal to n; 6, an interactive multi-model filtering algorithm is performed in the state when k is larger than or equal to 3 and smaller than or equal to n; 7, when k is larger than n, generalized high-order interactive multi-model filtering is performed. The method is used for the field of maneuvering target tracking.

The invention belongs to the technical field of online atmospheric-pollutant monitoring, and discloses an online atmospheric-pollutant monitoring system based on a wireless cloud sensing network. Thesystem is provided with a user terminal, a data aggregation terminal, a wireless cloud sensing network, a system analysis module, a temperature monitoring module, a pollutant monitoring module, another temperature monitoring module and a time control module. The temperature monitoring module, the pollutant monitoring module and the another temperature monitoring module are connected with the timecontrol module and the system analysis module through wires. The system analysis module transmits data to the wireless cloud sensing network through a GTiBee protocol. The data aggregation terminal receives and aggregates the data of the wireless cloud sensing network. The user terminal knows an atmospheric pollution situation through accessing the data aggregation terminal, and controls monitoring equipment. According to the system, atmospheric pollutants and atmospheric temperature and humidity can be automatically monitored online, the data are transmitted to the data aggregation terminal through the wireless cloud sensing network, and finally, observation is carried out through the user terminal, and control is carried out.

The embodiment of the invention provides a method for compressing streaming data. The method comprises the following steps: constructing a plurality of segments according to acquired streaming data and a predefined maximum error; determining a target piecewise linear function according to the plurality of segments, wherein the target piecewise linear function comprises a plurality of linear functions, and the intersection of a value range of independent variables of every two linear functions in the plurality of linear functions at most includes one value; and outputting a reference data point according to the target piecewise linear function, wherein the reference data point comprises a continuous point and a discontinuous point of the target piecewise linear function. Therefore, in the embodiment of the invention, the plurality of segments are constructed according to a plurality of data points and the maximum error, further the target piecewise linear function is determined according to the plurality of segments, and the continuous point and the discontinuous point of the target piecewise linear function are used for representing compressed streaming data. The method of the embodiment of the invention can guarantee that the target piecewise linear function has a minimum number of parameters, so that the requirement for a storage space is the lowest.

The invention discloses a method for generating a query-oriented video abstract by using a convolutional multilayer attention network mechanism. The method comprises the following steps: 1) cutting agroup of videos to construct video clips; and extracting visual features of each lens of the video by using a full convolutional neural network; 2) learning semantic relationships among all shots in the video clip by using a local self-attention mechanism, and generating visual features of the video shots; 3) learning a semantic relationship between different segments of the video by utilizing a query-related global attention mechanism, and generating query-oriented video shot visual features; 4) calculating a similarity score between the video shot and the user query to generate a query-related video abstract. Compared with a common video abstract solution, the method has the advantages that a convolutional multi-layer attention mechanism is utilized, video visual features related to query can be reflected more accurately, and a more consistent video abstract is generated. The effect obtained in the video abstraction is better than that obtained in a traditional method.

The invention belongs to the technical field of precipitation estimation, and discloses a cloud block characteristic change-based precipitation estimation method of a geostationary meteorological satellite. The method includes: acquiring a precipitation simulation parameter characteristic set of the satellite to depict an occurrence and development process of cloud precipitation, and selecting and using an extremum normalization method to normalize different-dimensional cloud image characteristic parameters; and constructing a three-layer forward-type backpropagation neural-network-based precipitation estimation model, which is used for precipitation estimation of a region, of the satellite, and adopting a multi-index system to analyze the precipitation simulation precision of the model. According to the method, the deficiency of conventional meteorological observation can be largely compensated by precipitation distribution derived by utilizing remote sensing data of the meteorological satellite, and richer precipitation information can be provided; and short-term near forecasting is facilitated, monitoring flood disasters is facilitated, early warning for geological disasters is facilitated, and the method has important significance for accuracy rate increasing of weather forecasting and disaster prevention and mitigation.

The invention discloses an optical remote sensing image segmentation method based on multi-scale lightweight hole convolution, and mainly solves the problems of large storage space occupied by a network and poor image segmentation effect in the prior art. The method comprises the steps of acquiring optical remote sensing image data, and dividing a training sample set and a test sample set; constructing a multi-scale lightweight hole convolution network formed by cascading a feature extraction lower sampling sub-network, a bottom layer sub-network and an image recovery upper sampling sub-network; training the constructed multi-scale lightweight hole convolutional network by using the training sample set; and inputting the test sample set into the trained multi-scale lightweight hole convolutional network for testing to obtain a segmentation result of the optical remote sensing image. According to the method, the storage space occupied by the segmentation network is reduced, the segmentation precision of the optical remote sensing image is improved, and the method can be used for land planning management, vegetation resource investigation and environment monitoring.

The invention discloses a method and device for backing off contention type slot time of a media access control layer. According to the scheme, the method adopted by the device is achieved based on a BC and a BSC, the device is applied to nodes participating in the contention in a CBTS, and after the contention nodes successfully transmit a frame inside a CW of the current CBTS, the BSC is subtracted by X progressively, wherein X is a positive integer. When the contention nodes detect that other nodes in the CW of the current CBTS transmit a frame, the BSC is subtracted by Y progressively, wherein Y is the positive integer. By means of the method and device, the sending opportunities can be increased in the subsequent contention for nodes not having the opportunity to be sent, and meanwhile the number of parameters in the current back-off method is reduced.

The invention discloses an improved target detection method based on a Faster RCNN algorithm, and the method comprises the following steps: (1) inputting a picture, carrying out the preprocessing of the input picture, and adjusting the size of the input picture; (2) performing feature extraction through a DenseNet network, and then dividing an extracted feature map into an upper branch and a lowerbranch; (3) generating a candidate box for the upper branch obtained in the step (2) through an RPN network, and obtaining a proposal feature map with a fixed size by a lower branch ROI network through a feature map generated by a DenseNet network and the candidate box generated by the upper branch; and (4) performing target classification and positioning through a prediction module.

The invention discloses a method for inverting rock pore distribution characteristics by utilizing a pore and fracture medium elastic wave theory. The method comprises the following processing steps of: 1, measuring saturation and drying speeds of a rock core under different pressures, and determining properties such as rock density; 2, simulating an elastic wave velocity of a rock by using a porefracture theory of polymorphic fractures; 3, calculating a pore aspect ratio spectrum under different pressure points according to the pore aspect ratio spectrum under the effective pressure of 0 forfractures with various forms in the rock; 4, establishing an inversion objective function; 5, setting fractures with various aspect ratios, wherein the fractures comprise pores; and 6, repeatedly adjusting the aspect ratio and the fracture density of each form of fracture of the rock under the effective stress of 0 to enable the objective function to reach the minimum value, so as to obtain the pore aspect ratio spectrum of the rock under each pressure point. The method can obtain rock pore structure characteristics more accurately, and analyze the mechanical, acoustic and fluid permeabilityproperties of the rock.

The invention discloses a model-driven Turbo code deep learning decoding method, which comprises the following steps of: firstly, unfolding a Turbo code iterative decoding structure into a tiled structure, and replacing each iteration with a DNN decoding unit to form a network Turbo Net for Turbo code decoding; then constructing a graphic structure of a traditional Max-Log-MAP algorithm and carrying out parameterization of the graphic structure to obtain a deep neural network based on the Max-Log-MAP algorithm as a sub-network in a TurboNet decoding unit, so that an SISO decoder in a traditional decoding structure and calculation of external information obtained through output of the SISO decoder are replaced; training a TurboNet composed of M DNN units to obtain model parameters; and finally, normalizing an output value of the Turbo Net by using a sigmoid function, and performing hard decision on a normalization result to obtain an estimated value of the real information sequence u todecode the Turbo code. According to the invention, Max-can be improved; log Compared with a model driven by pure data, the error rate performance of the Max-Log-MAP algorithm is improved, the numberof parameters is reduced by two orders of magnitudes, and the time delay is greatly reduced.

The invention relates to an improved YOLOv4-based shielded pedestrian real-time detection method, and belongs to the field of computer vision. The method comprises the following steps: acquiring a data set and processing the data set; clustering is carried out by using a Kmeans + + algorithm, and a final prior frame is generated; feature extraction is carried out by using a trunk feature network fusion channel attention mechanism; utilizing a spatial pyramid SPP module to carry out maximum pooling and merging on feature maps extracted by the backbone feature network; performing feature fusion processing on the last four layers of feature maps obtained by the large residual block in the backbone feature network; performing result prediction on the feature map after feature fusion through a YOLO detection head, predicting the target position and category, and training a model by using a loss function; and applying the optimal weight generated by model training to the model, and putting the test set pictures into the model for testing. The method realizes real-time high-precision detection of sheltered pedestrians, uses few parameters, and is high in detection speed.

The invention discloses a deep learning-based complex road condition perception system and apparatus. The deep learning-based complex road condition perception system comprises the five steps of 1, collecting data of an image sensor; 2, preprocessing the image data, wherein the preprocessing comprises white balance, gamma correction and denoising processing; 3, performing feature extraction by using a first deep neural network; 4, performing target locating by using features extracted by a second deep neural network; and 5, performing target identification on a locating region by using a thirddeep neural network to obtain a final detection result. The apparatus is used for the deep learning-based complex road condition perception system.