Integration method for depth feature and traditional feature based on AdaRank

Abstract

The invention relates to an integration method for depth feature and traditional feature based on AdaRank. The main technical characteristics comprise: dividing image data, respectively establishing aimed at different parts and training a depth convolution and a neural network, used to obtain depth features; extracting traditional features from pedestrian re-identification data, including LOMO features, ELF6 features, and Hog3D features; selecting the following three metric learning methods, KISSME, kLFDA, and LMNN; all the features and the three metric learning methods being combined and spanned to a Cartesian product, to obtain a series of weak sorters; using an AdaRank algorithm, performing ensemble learning on the weak sorters, to finally obtain a strong sorter. The method is reasonable in design, and combines depth learning, multi-feature, metric learning, and ensemble learning, and learns in an integrated manner through establishing the weak sorters, so integrated performance of a system is much better that using a single feature and a single metric learning, system integrated matching ratio is greatly improved, and good performance is obtained.

Description

technical field [0001] The invention belongs to the technical field of computer vision recognition, in particular to an integration method of deep features and traditional features based on AdaRank. Background technique [0002] As the scope of monitoring increases, the monitoring data shows explosive growth. It is obviously very inefficient to rely on human eyes to identify pedestrians in the monitoring screen. The task of pedestrian re-identification technology is to rely on computer vision technology to solve the problem of pedestrian identity matching in non-overlapping monitoring fields of view. [0003] Existing person re-identification algorithms are mainly divided into two categories, one is the traditional method, which consists of two steps of feature extraction and metric learning. In the feature extraction stage, the algorithm mines useful information according to the characteristics of the data and organizes it into features, which need to be descriptive, disti...

AI Technical Summary

Problems solved by technology

It is obviously very inefficient to rely on human eyes to identify the identity of pedestrians in the monitoring screen. The task of pedestrian re-identification technology is to rely on computer vision technology to solve the problem of pedestrian identity matching in non-overlapping monitoring fields of view.

Images