Multi-target tracking method based on cross-task mutual learning, terminal equipment and medium

Abstract

The invention discloses a multi-target tracking method based on cross-task mutual learning. The method comprises the following steps: S1, reading a current frame RGB image of a video sequence; s2, extracting universal features of the RGB image of the current frame by using a backbone network; s3, performing feature decoupling on the general features of the current frame of RGB image to obtain detection alignment embedding and re-identification alignment embedding; s4, obtaining a detection basic output according to the detection alignment embedding; s5, a cross direction Transform is used for obtaining re-identification output according to re-identification alignment embedding; s6, performing cross-task interactive learning on the detection basic output and the re-identification output to obtain the detection basic output and the re-identification output after interaction; s7, obtaining a target position frame on the current frame according to the interacted detection basic output, and querying a re-identification embedding on the interacted re-identification output according to the position of the central point; and S8, carrying out inter-frame association on the current frame and the previous frame according to re-identification embedding. According to the invention, efficient multi-target tracking can be realized in a complex scene.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a multi-target tracking method based on cross-task mutual learning. Background technique [0002] With the extensive theoretical research on computer vision and the business needs in practical scenarios, object tracking, especially multi-object tracking, has gradually become an important field. Due to the complexity of real scenes and the variability of pedestrian motion, there are still many problems to be solved in pedestrian multi-target tracking. At present, multi-target tracking is mainly divided into two-stage methods that first detect targets and then extract target features, and single-stage methods that simultaneously extract detection and re-identification features. For a long time when multi-task joint learning has not been popular, most research on multi-target tracking is two-stage, but both stages of the two-stage method need to use deep neural networks to ...

AI Technical Summary

Problems solved by technology

However, this joint network still faces many problems: (1) The features required for detection and re-identification are actually misaligned. The former needs boundary-aware information while the latter needs identity-aware information. How to ensure that the two It is very important for each task to obtain their own demand information; (2) For the re-identification task, the information of other targets and the environment around the target is crucial to enhance the discrimination of the target itself. The recognition task is very critical; (3) detection and re-identification can complement each other at the task level, how to achieve this complementarity is the main problem in building a more efficient multi-branch model

Images