Pedestrian retrieval method for carrying out multi-feature fusion on the basis of neural network

Abstract

The invention relates to a video analysis technology, in particular to a pedestrian retrieval method for carrying out multi-feature fusion on the basis of a neural network. Pedestrian calculation features detected in a video are stored into a feature database of a set of pedestrians to be detected, and then, features are calculated for a pedestrian to be detected and are compared with a feature database to obtain a retrieval result which ranks top and is high in similarity. Through the calculation of various retrieval features and feature distances, an optimal distance weight W is used for carrying out integration, so that the retrieval result is similar to the upper part of the body and the lower part of the body of the pedestrian to be inquired, and one feature distance is used for carrying out sorting to improve retrieval convenience. The method has the characteristics of wide applicable range, high accuracy and convenience in application. By use of the method, the problems of low accuracy and retrieval similarity in pedestrian detection based on a surveillance video are solved, and the problems that various detection methods and feature distances can not be favorably combined, the applicable range is narrow and application is complex can be overcome.

Description

technical field [0001] The invention relates to a video analysis technology, in particular to a pedestrian retrieval method based on neural network for multi-feature fusion. Background technique [0002] At present, there are many methods for image retrieval and pedestrian retrieval, such as image retrieval method CEDD (CEDD: Color and Edge Directivity Descriptor. A Compact Descriptor for Image Indexing and Retrieval, Savvas A.2008), pedestrian retrieval method WHOS (Person Re-Identification by Iterative Re- Weighted Sparse Ranking, Giuseppe Lisanti, 2015), these methods can achieve better retrieval results for some scientific research datasets, such as the pedestrian retrieval dataset ViPeR Dataset (https: / / vision.soe.ucsc.edu / node / 178) , but for the pedestrians in the actual surveillance video, the retrieval is not ideal, and it needs to be integrated to form a new retrieval feature. [0003] From the perspective of retrieval results, some methods, such as WHOS, although ...

AI Technical Summary

Problems solved by technology

[0002] At present, there are many methods for image retrieval and pedestrian retrieval, such as image retrieval method CEDD (CEDD: Color and Edge Directivity Descriptor. A Compact Descriptor for Image Indexing and Retrieval, Savvas A.2008), pedestrian retrieval method WHOS (Person Re-Identification by Iterative Re- Weighted Sparse Ranking, Giuseppe Lisanti, 2015), these methods can achieve better retrieval results for some scientific research datasets, such as the pedestrian retrieval dataset ViPeR Dataset (https: / / vision.soe.ucsc.edu / node / 178) , but for the pedestrians in the actual surveillance video, the retrieval is not ideal, and it needs to be integrated to form a new retrieval feature

[0003] From the perspective of retrieval results, some methods, such as WHOS, although the retrieval results include the pedestrians to be queried, that is, the retrieval is successful, but the similarity between other pedestrians in the retrieval results and the pedestrians to be queried is not high, and cannot provide users with more references For example, if the pedestrians to be queried have blue upper body and black trousers lower body, many of the pedestrians who rank first in the search results are not “blue upper body and black lower body”; while others can be retrieved by body parts Methods, such as: AGeneral Method for Appearance-based People Search Based on Textual Queries, R. Satta, 2012, the similarity between the retrieval results and the pedestrians to be queried is relatively large, but the retrieval process and feature distance are relatively complicated, and require more explicit use One feature distance and retrieval filtering link, cannot use one feature distance to represent the similarity

Images