Pedestrian re-recognition method based on depth-learning joint optimization

A pedestrian re-identification and deep learning technology, which is applied in the field of pedestrian re-identification based on deep learning joint optimization, can solve the problems of pedestrian re-identification performance degradation, feature discrimination degradation, and differences, achieving superior performance, enhancing detection capabilities, and solving problems. complex background effects

CN109102025AInactive Publication Date: 2018-12-28UNIV OF ELECTRONICS SCI & TECH OF CHINA
2 Cites 31 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Publication Date
2018-12-28
Estimated Expiration
Not applicable · inactive patent

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

The invention discloses a pedestrian re-identification method based on depth learning joint optimization, which comprises the following steps: 1, collecting and screening positive and negative pedestrian sample pairs with balanced quantity to construct a data set; 2, constructing a deep-learning Siamese neural network structure model, comprising a two-way front-end convolution neural network and amulti-level feature fusion module, input positive and negative pedestrian samples into that model, and extracting the Hyper features of two different pedestrian; 3, sending the Hyper features of twodifferent pedestrian into a classification network and a verification network, combining the classification network and the verification network, combining the classification los function and the verification loss function, and optimizing the parameters of the neural network structure model. In the method, the depth convolution neural network and HyperNet network are combined to extract multi-scale features to enhance the detection ability of pedestrian target, and the verification model and classification model are combined to optimize the network structure, and the excellent pedestrian re-recognition neural network structure model is obtained.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The invention belongs to the technical field of pedestrian re-identification in computer vision, and in particular relates to a pedestrian re-identification method based on deep learning joint optimization. Background technique

[0002] Pedestrian re-identification is one of the important topics in the field of computer vision and pattern recognition. Pedestrian re-identification refers to retrieving a given pedestrian target in multiple cameras, and correlating and matching the retrieval results to quickly and accurately find the target pedestrian. Live footage and tracks under multiple cameras. Because of its important significance in the fields of intelligent video surveillance and multi-target tracking, it has attracted more and more attention from scientific researchers in related fields, governments and public security departments in recent years.

[0003] Person re-identification mainly studies the use of visual features to match pedestrian obj...

Examples

Embodiment 1

[0063] A preferred embodiment of the present invention provides a pedestrian re-identification method based on deep learning joint optimization, the method steps are:

[0064] Step 1. Collect and screen a balanced number of positive and negative pedestrian sample pairs to construct a data set. Specifically:

[0065] Step 1.1, randomly select the input pedestrian sample pictures through the online sampling layer;

[0066] Step 1.2. Filter the corresponding labels of positive / negative sample pairs in the pedestrian sample pictures, and obtain a data set composed of positive and negative pedestrian sample pairs with a balanced number.

[0067] figure 1It shows the process of collecting sample pairs in this method, including the flow chart of selecting the input picture of the online sampling layer and the flow chart of screening positive / negative sample pairs corresponding to the label pairlabel. The datasets used for network structure training include Market-1501 and CUHK-SYS...