Pedestrian re-identification feature descriptor based on multi-task learning

Abstract

The invention discloses a pedestrian re-identification feature descriptor based on multi-task learning. A twin network structure with paired input is adopted; the method comprises the following stepsof: sending local maximum occurrence (LOMO) features and deep features into a network together, and mapping the features into a single feature space for training to form a new TDFN (Time Division Multiplexing Network) model by using a time division duplex (TDFN) algorithm, so that a new TDFN (Time Division Multiplexing Network) is formed. The neural network self-learning characteristic is utilized, and the network is updated by combining the loss functions of multiple tasks, so that the deep features learn more detail information complementary with the manual local features, and new features with higher discriminability are obtained. Experiments show that the average precision mAP and Rank-1 precision of the new features of the method are superior to those of a global descriptor directly extracted from a twin network. The method is suitable for application systems in the aspects of safety and monitoring, such as video monitoring analysis and content-based image and video retrieval.

Description

technical field [0001] The invention relates to the pedestrian re-identification problem in the field of video intelligent monitoring, in particular to a pedestrian re-identification feature descriptor based on multi-task learning and a new network model TDFN (Traditional and Deep features Fusion Network). Background technique [0002] Pedestrian re-identification (Re-Identification) aims to match image frames containing the same pedestrian in cross-camera surveillance video, which is a challenging topic in the field of intelligent surveillance analysis. Due to its important applications in security and surveillance, such as video surveillance analysis and content-based image and video retrieval, person re-ID has attracted extensive attention in both industry and academia. Re-identification models usually include two parts: representation learning and metric learning. In typical re-identification, a single feature is usually used to describe each pedestrian image, and then ...

Images