34results about How to "Improve re-identification performance" patented technology

The invention belongs to the technical field of person recognition, and discloses a method, system and terminal for re-recognition of people in a video-based smog scene, constructing a symmetrical non-local codec K estimation network model to perform defogging processing on the video; constructing a discriminant network model, And based on the built discriminant network model, it is estimated whether the input video is a normal video or a haze-free video generated by the dehazing sub-network; a non-local double-attention person re-identification sub-network model is constructed to re-identify the person. The invention can solve the problem of difficult re-recognition caused by fog in the re-recognition of characters in the video. The present invention can well complete person re-identification in foggy video. The entire process of the present invention is an end-to-end design, which can be used more simply. The invention is a video-based person re-identification technology in a smoky scene, which can complete end-to-end discrimination and can well complete person re-identification.

The invention discloses a double-angle sequencing optimization based pedestrian re-identification method, and belongs to the technical field of surveillance video retrieval. The double-angle sequencing optimization based pedestrian re-identification method comprises the steps of: firstly obtaining initial sequencing results through ways based on global characteristics and local characteristics; then taking an intersection set GKc(p) of front k results of the two sequencing results as initial values for inquiring a more reliable nearest neighbor; then regarding each goal gKc(j) in GKc(p) as a new inquiring factor for crossed opposite inquiry, weighting and fusing a sequence list of the opposite inquiry to obtain a whole optimized sequence list; finally, establishing a double-layer map for each gKc(j) in the GKc(p), and calculating neighbor similarity of the map to perform further rearrangement on the GKc(p) so as to further obtain a better re-identification effect.

The invention discloses a pedestrian re-identification method based on synchronous enhancement of video appearance and motion information. During training, two modules, an appearance enhancement module AEM and a motion enhancement module MEM, respectively enhance the pedestrian appearance and motion information in a backbone network. The appearance enhancement module AEM uses the attribute recognition model trained by the existing large-scale pedestrian attribute datasets to provide attribute pseudo-labels for large-scale pedestrian video datasets, and enhances appearance and semantic information through attribute learning; the motion enhancement module MEM uses video The prediction model predicts the gait information of pedestrians, enhances the gait information features with identity discrimination in the backbone network of pedestrian feature extraction, and improves the performance of pedestrian re-identification. In practical applications, only the backbone network for pedestrian feature extraction needs to be retained, and higher pedestrian re-identification performance can be obtained without increasing network complexity and model size. The enhanced backbone features achieve higher accuracy in video-based person re-identification tasks.

The invention discloses a pedestrian re-identification method for solving component misalignment. That is to construct a deep representation of pedestrian images, extract multi-layer features through the backbone network model, use sub-modules to enhance and fuse multi-layer features, form a multi-branch structure, extract component features and global features of each branch; train the constructed network model, Define the configuration related to the experiment, optimize the model parameters of the network model; pedestrian re-identification, extract the deep representation of the query image through the trained network model, and use the second paradigm to normalize according to the similarity between each query image and the query set, Returns the recognition results for each query image. The best person re-identification performance at this stage is achieved by the person re-identification method based on the fusion of multi-scale features to solve the misalignment of parts.

The present invention provides a cross-camera re-identification fusion method for objects with similar appearances, which uses a deep convolutional neural network to extract the global feature map of the picture, and extracts the appearance vector of the target on the global feature map according to the target detection result; encodes the camera, Generate a view vector containing observation view angle information; generate a position vector of the target according to the position of the detection frame corresponding to the target in the image coordinate system. The three vectors are fused and transformed to generate the target representation vector. The network is trained by optimizing the triplet loss function, and the representation vector for re-identification is learned. During the training process, a combination of offline mining and online mining is used to generate and update the triplet dataset. Finally, the hierarchical clustering algorithm with constraints is used to cluster the representation vectors corresponding to the targets in different cameras to realize cross-camera target re-identification. At the same time, a cross-camera re-identification fusion system for objects with similar appearance is provided. The invention improves the accuracy of re-identification.

The invention discloses a pedestrian re-identification method based on a global distance scale loss function. The specific steps include: performing data enhancement processing on the training data of the pedestrian re-identification data set; randomly selecting each batch of data; constructing a human body component-based Deep neural network and initialize the network; use the cross-entropy loss function and the loss function based on the global distance scale to supervise the training of the deep neural network at the same time; perform feature extraction on the target pedestrian image and pedestrian images in the pedestrian image library, and perform cosine similarity Degree calculation and sorting to get the recognition results. The present invention proposes a loss function of the global distance scale based on statistical characteristics, which can effectively avoid noise interference and reduce the risk of over-fitting, and improve model robustness and generalization ability.

The invention discloses a method for generating pedestrian re-identification data under different illumination conditions based on an adversarial network. The method comprises the following steps: firstly, acquiring videos of different time periods in a monitoring video, namely monitoring videos under different illumination conditions; Extracting representative monitoring images from the acquiredmonitoring videos, and marking the representative monitoring images into three different categories of'too dark, normal and too bright 'according to illumination conditions; Sending the images with the illumination condition labeling information into a generative adversarial network for training, wherein the trained generative network is used for generating data under different illumination conditions; And generating the pedestrian re-identification data into data under different illumination conditions through the trained generation network; And finally, adding the obtained data as a data augmentation form into pedestrian re-identification model training, and at the moment, enabling the model to identify the characteristics of the same target under different illumination conditions, i.e.,enhancing the identification performance of the model on the target under different illumination conditions.

The invention discloses a pedestrian re-identification method based on label uncertainty and a human body component model. The pedestrian re-identification method comprises the following steps: (1) constructing a deep neural network model based on a human body component; (2) initializing the constructed deep neural network model, and performing training according to the constructed deep neural network structure to obtain another model; (3) training the initialized deep neural network model and updating parameters in the network; (4) carrying out feature extraction on the target pedestrian image and pedestrian images in a pedestrian image library by adopting the trained deep neural network; And (5) performing cosine similarity calculation and sorting on the extracted features to obtain a recognition result. According to the method, the problems that local component information is not completely accurate and classification confidence coefficients of all local components are large in difference are effectively processed, the correct rate of pedestrian re-identification can be effectively improved, and / or the false identification rate can be effectively reduced.