45results about How to "Improve re-identification accuracy" patented technology

The invention belongs to the technical field of image processing, and discloses a license-plate-information-free vehicle re-identification method and system, a medium and a video monitoring system, and the license-plate-information-free vehicle re-identification method comprises the steps: obtaining a data set, and carrying out the data set division and preprocessing; taking ResNet-50 which is trained in advance as a backbone network, and extracting a basic feature tensor; adding a channel attention mechanism and a space attention mechanism to obtain a new feature tensor; designing four relatively independent networks for extracting higher-level semantic features respectively; utilizing the cross entropy loss function and the triple loss function to train and optimize the vehicle overall network to obtain a training model; testing the test image by using the trained model to obtain a re-identification initial sorting result; sorting the initial sorting result again by using a resortingalgorithm; and visualizing the final sorting result. According to the license-plate-information-free vehicle re-identification method, the vehicle re-identification rate under the condition of no license plate information is effectively improved, and the vehicle re-identification accuracy based on no license plate information under a complex scene is improved.

The invention relates to a method for re-identifying persons on the basis of deep learning encoding models. The method includes steps of firstly, encoding initial SIFT features in bottom-up modes by the aid of unsupervised RBM (restricted Boltzmann machine) networks to obtain visual dictionaries; secondly, carrying out supervised fine adjustment on integral network parameters in top-down modes; thirdly, carrying out supervised fine adjustment on the initial visual dictionaries by the aid of error back propagation and acquiring new image expression modes, namely, image deep learning representation vectors, of video images; fourthly, training linear SVM (support vector machine) classifiers by the aid of the image deep learning representation vectors so as to classify and identify pedestrians. The method has the advantages that the problems of poor effects and low robustness due to poor surveillance video quality and viewing angle and illumination difference of the traditional technologies for extracting features and the problem of high computational complexity of the traditional classifiers can be effectively solved by the aid of the method; the person target detection accuracy and the feature expression performance can be effectively improved, and the pedestrians in surveillance video can be efficiently identified.

The invention relates to the field of the identification method, and particularly relates to a personnel reidentification method based on deep learning and distance metric learning. The identificationmethod comprises the steps that (1) a pedestrian target detection method based on the convolutional neural network is adopted to process the video data so as to detect the pedestrian target in the video; (2) the initial characteristics of the pedestrian target are coded by using an unsupervised RBM network through the bottom-up mode so as to obtain a visual dictionary having sparsity and selectivity; (3) supervised fine adjustment is performed on the initial visual dictionary by using error back propagation so as to obtain the new image expression mode of the video image, i.e. the image deeplearning representation vector; and (4) the metric space closer to the real semantics is acquired by using the distance metric learning method of characteristic grouping and characteristic value optimization, and the pedestrian target is identified by using a linear SVM classifier. The essential attributes of the image can be more accurately expressed so as to greatly enhance the accuracy of pedestrian reidentification.

The invention provides a pedestrian re-identification system and method, and belongs to the technical field of pedestrian re-identification. The system comprises the following steps of carries out image reconstruction on different original images through a deep learning network based on sparse coding to obtain corresponding reconstruction matrixes; extracting a feature vector in each reconstruction matrix in combination with an attention mechanism; calculating a classification loss result and a verification loss result of the feature vector; and judging whether the feature extraction module converges or not according to the classification loss result and the verification loss result, if yes, calculating a difference degree between the feature vectors of different reconstruction matrixes, if the difference degree is greater than a set threshold value, determining that the feature vectors do not belong to the same pedestrian, and if the difference degree is less than the set threshold value, determining that the feature vectors belong to the same pedestrian. According to the invention, the reconstructed sub-network is used to reconstruct the image so as to improve the image definition, and the multi-task loss function is used to shorten the distance between the same individuals, thereby improving the feature representation capability and discrimination capability of the network,and improving the pedestrian re-identification accuracy.

The invention discloses a resolution adaptive feature extracting and fusing forpedestrian re-identification method. According to the invention, biological features and appearance features are fused, and a method fusing face features and experience features is used for identifying pedestrians, so feature diversity is improved; according to performance of the features in different image scales, different pedestrianfeatures are compared in different scales; a screening mechanism is adopted in which fusion features obtained by use of color features and profile features are firstly screened, then the face features are used for complementing screening results and at last texture features are extracted from the screened pedestrians, so color region features and weighing color features are greatly stabilized; and by using overall features and local features obtained via fusion and extraction by use of the self-adaptive weighing method, the fusion features are obtained. In this way, via self-adaptive fusion of the biological features and the appearance features, accuracy of the method can be improved; and by extracting the appearance features in a low scale, via the screening mechanism, complexity can be reduced.

The pedestrian re-identification system comprises a multi-stream feature distance fusion system used for calculating the image similarity between an image p to be detected and each reference image inan initial reference image set G, a sorting system used for sorting the reference images according to the image similarity, and a re-sorting system based on k neighborhood distribution scores. And thereordering system is used for reordering the reference images in the initial ordering list according to the k neighborhood distribution score on the basis of the initial ordering list. The inventionfurther discloses a pedestrian re-identification method, and the pedestrian re-identification system is adopted by the pedestrian re-identification method. The pedestrian re-identification method aimsat two main factors influencing pedestrian re-identification, namely image similarity calculation and image sorting, improvement is carried out respectively, and compared with the prior art, the improvement of any one aspect can improve the accuracy of pedestrian re-identification. And an optimal pedestrian re-identification effect can be obtained through the combination of the improvement of thetwo aspects.

A pedestrian re-identification method based on a skeleton posture comprises the following steps: acquiring a pedestrian image, and constructing a target pedestrian template; collecting image frames and sequentially segmenting all pedestrian images under various postures in the image through skeleton detection; performing size standardization on all segmented pedestrian images and skeleton information according to the target pedestrian skeleton template; according to the pedestrian skeleton information, sequentially carrying out local feature region segmentation on all the pedestrian images, and carrying out tilt correction on the local feature regions of all the pedestrians to obtain a local feature image set consistent with the target pedestrian template attitude; and establishing a target pedestrian local feature fusion recognition model, and comprehensively calculating the similarity between the target pedestrian template and all pedestrians detected in real time to realize accuraterecognition of the target pedestrian. According to the invention, effective pedestrian identification under the influence of abnormal postures such as bowing, limb inclination and the like in a multi-person scene in a complex environment can be realized, and the pedestrian identification accuracy is improved.

The invention discloses a multi-attribute depth characteristic-based vehicle re-recognition method. The method comprises the steps of extracting, by using a characteristic extraction model, a depth characteristic of a tested image set of an Ath pooling layer, obtaining, by using the depth characteristic of the tested image set and a W matrix, a Mahalanobis distance between a depth characteristic of a tested image in a searched set and a depth characteristic of a target image in a candidate set, ranking in an ascending order according to the Mahalanobis distance so as to obtain a similarity ranking result of the tested image and the target image, wherein tested image sets comprise a seeking set and a searching set, the tested image sets refer to images comprising vehicles, and the characteristic extraction model is trained through steps of accessing a vehicle multi-attribute classifier behind the Ath pooling layer of GoogLeNet, so as to obtain improved GoogLeNet, training by using training images to improve GoogLeNet so as to obtain the characteristic extraction model. The method simplifies the model training process and greatly improves re-recognition accuracy, and the model has strong generalization performance.

The invention provides an unsupervised pedestrian re-identification method based on pedestrian attribute adaptive learning. The method comprises the following steps: firstly, pre-training an adjustedResNet50 network by utilizing a source video set with a label to obtain a pre-trained pedestrian attribute CNN model; secondly, space-time constraint information of the target video set is fully utilized on the target video set without any mark to generate a training sample on line; then, inputting the generated training sample into a Triplet network, and carrying out parameter fine tuning on a pre-trained pedestrian CNN model to obtain pedestrian features with higher discrimination; and finally, performing pedestrian feature extraction and similarity measurement by using the trained network to realize pedestrian re-identification. Characteristics learned by the method have higher robustness and discrimination, the accuracy of pedestrian re-identification can be remarkably improved, and the application range of pedestrian re-identification is expanded.

The invention discloses a pedestrian re-identification method based on self-excitation discriminative feature learning, which comprises the following steps: (1) selecting a pedestrian re-identification network, and adding a negative branch on the original network; (2) in the training stage, generating a classification loss function by the original network, generating a confrontation loss functionand a mutual exclusion response item between the original network and the negative branch to form an objective function, and utilizing a random gradient descent method to optimize the whole network; (3) in the test stage, removing a negative branch, only keeping the part, in front of the classifier, of the original network as a trained network model, and inputting a pedestrian picture for extracting a feature vector test; and (4) in the pedestrian retrieval stage, extracting the feature vector of each picture in the picture library by using the trained network model, and selecting the identityof the picture with the highest similarity with the feature vector of the to-be-queried pedestrian picture as a final recognition result. According to the invention, the effect of the existing pedestrian re-identification network can be improved.

The invention discloses a deep Hash pedestrian re-identification method, which comprises the following steps of 1, constructing a deep neural network, wherein the deep neural network comprises a feature learning module and a Hash learning module, the feature learning module adopts a Resnet network, and the Hash learning module comprises a full connection layer and a tanh function layer; 2, training the deep neural network, 1) preparing the pedestrian pictures, and 2) sending the training pictures into the deep neural network for training, including feature learning, Hash learning and loss function learning; 3) carrying out the network optimization and parameter updating; step 3, testing the deep neural network, obtaining a loose hash code through the feature learning module and the hash learning module, and then converting the loose hash code into the strict -1 and 1 codes through a symbol function, and calculating the Euclidean distance between the Hash codes corresponding to the pedestrian pictures in query and gallery to carry out feature matching. The method has the advantages that the pedestrian re-identification accuracy is improved, and the pedestrian re-identification timeis shortened.

The invention discloses a video pedestrian re-identification method based on region guidance and space-time attention, and the method comprises the steps: firstly constructing global features, calculating the global features extracted by each frame in a video stream based on a time attention model, and carrying out aggregation to obtain a global feature vector; constructing region features, horizontally dividing the extracted pedestrian depth feature map into four blocks, generating respective guide frames of the four regions through a key frame generation layer to extract the corresponding region features, and performing calculation in combination with a space-time attention model to obtain a region feature vector; wherein the feature vector of the to-be-identified pedestrian video streamis obtained by combining the global feature vector and the regional feature vector, and is compared with the feature vector of the pedestrian in the video image of the selected geographic region to obtain the pedestrian target video stream with the minimum distance and output a final re-identification matching result. The method can solve the problems of poor pedestrian video image imaging quality, serious detail loss, pedestrian part loss and low video pedestrian re-identification accuracy caused by feature extraction difficulty.

The invention provides an unsupervised pedestrian re-identification method based on category adaptive clustering. The method comprises the following steps: firstly, pre-distributing an initializationlabel for a label-free data set, and extracting high-dimensional features of a pedestrian image by utilizing a CNN model; then, carrying out staged category adaptive clustering on the feature set, updating an image label, calculating an average contour coefficient of a clustering result of the stage, and alternately updating a feature extraction CNN model and a category adaptive clustering model until an inflection point appears in the average contour coefficient; and finally, extracting test pedestrian image features by using the CNN model at the inflection point obtained by training, and calculating a feature distance to obtain an unsupervised pedestrian re-identification result. The method can solve the problem that the number of individuals in an unmarked data set pedestrian re-identification task cannot be determined, can dynamically adapt to data sets of different scales, and has high algorithm efficiency and precision.

The invention discloses a vehicle re-identification method based on double-subnetworks. The method comprises the following five modules: obtaining a vehicle image data set, preprocessing vehicle image data, constructing a vehicle image feature extraction neural network, training a vehicle re-identification model, extracting a to-be-identified vehicle image feature vector, and calculating the feature similarity between the to-be-identified vehicle image feature vector and a test set image. The vehicle re-identification method based on the double sub-networks aims to construct a neural network model with a strong discrimination capability, further solves the influence of factors such as illumination and a visual angle on a vehicle re-identification result, and improves the accuracy of vehicle re-identification. According to the invention, a double-subnetwork is designed on the basis of fusion of CNN and Transform structures, and effective global feature information can be obtained while multi-granularity local features of a vehicle image region are learned cooperatively, so that the discriminative characterization capability of a vehicle re-identification model is improved.

The invention discloses a pedestrian re-identification method and device based on space-time analysis and depth features. In pedestrian re-identification application, global search for pedestrian images in an actual large-scale video monitoring scene has complexity and irrationality, and in order to further improve the identification accuracy and the identification speed, the invention provides amethod combining spatio-temporal information analysis and depth feature extraction. Firstly, the moving speed of a pedestrian is obtained through analysis to accord with gamma distribution, and then the space-time information of the pedestrian is further analyzed through the distribution to obtain the space-time prior probability of the pedestrian; training a convolutional neural network on the large-scale data set in combination with an actually acquired image, and extracting depth features to calculate a visual space-time probability; and finally, combining the two probabilities to judge whether the two images are the same pedestrian or not. The pedestrian re-identification efficiency can be effectively improved from massive monitoring or collected data in an actual large-scale video monitoring application scene, the high pedestrian re-identification precision can be kept, and the efficient and accurate pedestrian re-identification effect is achieved.

The invention discloses a cross-modal pedestrian re-identification method based on adaptive pedestrian alignment. The method comprises the following steps: firstly, respectively extracting features of an infrared image and a visible light image by utilizing a multi-path network based on a residual network pre-training model ResNet50; then, linearly regressing a group of affine transformation parameters by utilizing the high-level characteristics of the visible light image to carry out adaptive affine transformation on the visible light image; after an aligned and corrected image is generated, extracting and fusing features of the image with features extracted from an original visible light image to serve as final features of the visible light image; and finally, mapping the features of the infrared image and the visible light image into the same feature space, and training in combination with an identity loss function and a most difficult batch sampling loss function to finally achieve higher recognition precision compared with a general cross-modal pedestrian re-recognition method. The method is mainly applied to a video monitoring intelligent analysis application system, and has wide application prospects in the fields of image retrieval, intelligent security and the like.

The invention discloses a pedestrian re-identification method and a pedestrian re-identification system based on ST-SSCA-Net. The method comprises the following steps of: constructing a network by using a pytorch; adopting a Yolov3 algorithm to position and track pedestrians in the video; cutting out a pedestrian range extracted by a Yolov3 algorithm, sending a picture of the pedestrian range intoa pedestrian re-identification algorithm based on ST-SSCA-Net to compare and identify the pedestrian identity of the pedestrian and the pedestrian identity of a picture library, and performing systemconstruction by an edge segment high-definition video camera and a real-time visualization system. Compared with a similar pedestrian re-identification algorithm, the method has the advantages that the acquisition of an attention mechanism is enhanced, and the model accuracy is improved. Meanwhile, the invention designs an enhanced triple loss for model training, and the clustering effect of themodel is greatly improved.

The invention discloses a target tracking method in a multi-camera scene, which realizes splicing of pictures of different cameras by combining YOLO-V4 with an improved DeepSort algorithm and an image splicing algorithm, and finally realizes multi-target tracking in a spliced video. In the aspect of data sets, a self-made intelligent trolley data set and a self-made vehicle re-identification data set containing an intelligent trolley are adopted. According to the vehicle re-identification method, multi-target tracking in a multi-camera scene is realized by constructing a rich data set, improving a model and splicing and fusing pictures, and the vehicle re-identification accuracy is well improved.

The invention provides a cross-modal pedestrian re-identification method based on double transformation alignment and blocking. The method comprises steps of firstly, a base branch network being used for extracting input infrared and visible light pedestrian image features, a group of affine transformation parameters being linearly regressed by using high-level features of the images, then an alignment image being generated by using the parameters, and the image being capable of effectively relieving modal difference of misalignment; next, horizontally dividing the aligned image into three blocks, then taking out the features of the three block images, and fusing the features with the aligned global features and original image features as the total features of the visible light and infrared images; next, the total features of the infrared and visible light images being mapped to the same embedding space; and finally, carrying out joint training by combining identity loss and the most difficult batch sampling loss function with the weight so as to improve the recognition precision. The method is mainly applied to a video monitoring intelligent analysis application system, and has wide application prospects in the fields of image retrieval, intelligent security and protection and the like.

The invention discloses a cross-domain pedestrian re-identification method and system fusing multiple source domains. The method comprises the steps that when a cross-domain pedestrian re-identification model is trained, multiple source domain data sets are obtained, the cross-domain pedestrian re-identification model is trained by training a representation learning network and a metric learning network through one source domain data set, the trained cross-domain pedestrian re-identification model is obtained, and when pedestrian re-identification is conducted, a pedestrian sample pair to be identified is inputted into the trained cross-domain pedestrian re-identification model to obtain a pedestrian re-identification result. By adopting the method and the system, the data of the pluralityof the source domains are fused in the training process, the feature representation of the pedestrian can be better learned, and compared with a single training model, the pedestrian re-identification accuracy can be improved, and the problem that the model performance is reduced due to inter-domain difference is effectively solved.