Target re-identification method based on local feature fusion

Abstract

The present invention discloses a target re-identification method based on local feature fusion. The method comprises: step S1, acquiring a foreground image of a to-be-processed image; step S2, performing area segmentation on a target in the foreground image by using a color segmentation operator and a body segmentation operator; step S3, extracting a color feature of a key area according to a weighted HSV color histogram, and extracting a body feature of the key area according to a body information descriptor; and step S4, according to similarity measurement of the color feature and the body feature, performing re-identification on the target. According to the target re-identification method based on local feature fusion, at least the technical problem how to enable one identified target to be still subjected to identification when the lighting condition, target position and posture, a scene and the like are changed and the identified target is shielded and re-emerges is solved; and the target re-identification method has the advantages that identification real-time performance is good and the technical guarantee is provided for application in the field of multi-target tracking, video monitoring and the like.

Description

technical field [0001] Embodiments of the present invention relate to the technical field of computer vision, and in particular, to a target re-identification method based on local feature fusion. Background technique [0002] The target re-identification method can be used in multi-target tracking, video surveillance and other fields, especially in the continuous tracking application of multiple targets in multiple scenarios under the framework of visual sensor network. [0003] At present, the research methods of target re-identification are mainly divided into two categories. The first method is to directly count and use the appearance characteristics of the target such as color and shape. The traditional color histogram does not contain spatial position information, which reduces the Similarly, when counting shape information, pixels farther from the center of gravity contain more shape information of the target, and this situation has not been given enough attention. T...

AI Technical Summary

Problems solved by technology

[0003] At present, the research methods of target re-identification are mainly divided into two categories. The first method is to directly count and use the appearance characteristics of the target such as color and shape. The traditional color histogram does not contain spatial position information, which reduces the similarly, when counting shape information, pixels farther from the center of gravity contain more shape information of the target, and this situation has not been given enough attention

The second type of method finds out the feature component with the largest degree of discrimination through training, establishes a similarity measurement model, and then realizes the re-identification of the target. In this type of method, it takes a long time to collect samples and adjust parameters, and the target type Retraining is required when replacing, and the versatility is low

In addition, in order to make the extracted features more robust, attention should be paid to key areas with stable features and rich information, and research in this area is still relatively scarce

[0004] The inventor found in the research that when a target that has been recognized reappears after the lighting conditions, target pose, scene, etc. change or is blocked, the existing technology cannot recognize it

Images