Pedestrian Image Retrieval Method Based on Cross-View Information and Quantization Error Coding

Abstract

The invention discloses a large-scale quick retrieval method of a pedestrian image on the basis of cross-horizon information and quantization error encoding. The large-scale quick retrieval method comprises the following steps: (1) establishing a combined model of the cross-horizon information of a pedestrian image and a quantization error; (2) solving an encoding hash function of a minimum cross-horizon quantization error; and (3) quickly matching cross-horizon pedestrian images. The large-scale quick retrieval method permits to solve the specific encoding hash function which aims at each horizon in the pedestrian image data of a large-scale overlap-free horizon so as to realize quick and accurate pedestrian image retrieval. The large-scale quick retrieval method permits each overlap-free horizon to solve the specific encoding hash function so as to improve the discrimination performance of the encoding hash function. Meanwhile, the large-scale quick retrieval method restrains a difference value between cross-horizon hash functions, and effectively avoids overfitting. Compared with a traditional supervision hash encoding and multi-modal hash encoding method, the large-scale quick retrieval method can obviously improve the retrieval accuracy of a large-scale pedestrian image open set test.

Description

technical field [0001] The invention relates to a fast retrieval method for large-scale pedestrian images in the field of video surveillance, in particular to a fast retrieval method for large-scale pedestrian images based on cross-view domain information and quantization error coding. Background technique [0002] Video surveillance is the guarantee of public security in public places. Government agencies and various business units often use non-overlapping cameras for video surveillance. With the increasing emphasis on public safety from all walks of life, the application of video surveillance is becoming more and more extensive, followed by a substantial increase in the size of video surveillance data. Therefore, the rapid retrieval of large-scale pedestrian images, especially without traffic Fast retrieval of cross-view large-scale pedestrian images with stacked cameras has become a new requirement. [0003] In order to quickly retrieve data, in recent years, there have...

AI Technical Summary

Problems solved by technology

[0003] In order to quickly retrieve data, in recent years, there have been many quantitative hash binary coding methods for data, such as local sensitive hashing, spectral hashing, supervised kernel hashing and other binary coding methods, which are mainly based on quantization error and supervision. Information encodes data, without considering the differences of data sources, and adopts a consistent encoding method for data, so it cannot solve the problems of illumination differences and background differences in pedestrian images of different cameras; in recent years, multi-modal data has appeared. The main strategy of this method is to maintain the correlation of different modal data of the same sample in the encoding process, but the use of supervisory information is not sufficient, which also leads to the low accuracy of this type of method in the application of pedestrian image retrieval

Images