An Improved End-to-End Multi-Target Tracking Method Based on Transformer Self-Attention

Abstract

Current multi-target multi-camera tracking algorithms demand increased requirements for re-identification accuracy and tracking reliability. This study proposed an improved end-to-end multi-target tracking algorithm that adapts to multi-view multi-scale scenes based on the self-attentive mechanism of the transformer’s encoder–decoder structure. A multi-dimensional feature extraction backbone network was combined with a self-built raster semantic map which was stored in the encoder for correlation and generated target position encoding and multi-dimensional feature vectors. The decoder incorporated four methods: spatial clustering and semantic filtering of multi-view targets; dynamic matching of multi-dimensional features; space–time logic-based multi-target tracking, and space–time convergence network (STCN)-based parameter passing. Through the fusion of multiple decoding methods, multi-camera targets were tracked in three dimensions: temporal logic, spatial logic, and feature matching. For the MOT17 dataset, this study’s method significantly outperformed the current state-of-the-art method by 2.2% on the multiple object tracking accuracy (MOTA) metric. Furthermore, this study proposed a retrospective mechanism for the first time and adopted a reverse-order processing method to optimize the historical mislabeled targets for improving the identification F1-score (IDF1). For the self-built dataset OVIT-MOT01, the IDF1 improved from 0.948 to 0.967, and the multi-camera tracking accuracy (MCTA) improved from 0.878 to 0.909, which significantly improved the continuous tracking accuracy and reliability.