Multi-Object Tracking Framework Based on Multi-Scale Temporal Feature Aggregation
Jialiang Liu, Xiaopeng Hu · 2023
Multi-object tracking (MOT) is a hot spot in computer vision field. There have been various related works in recent years. We propose a framework which is used for multi-object tracking tasks with multi-class in this paper. Our framework is based on one-shot method. Firstly, we propose a one-shot tracker. Secondly, we develop a frame attention module (FAM) to more effectively aggregate features of two adjacent frames. Thirdly, we introduce a convolutional block attention module for each head on each scale to decouple detection and association. Our framework based on an anchor-free detector. The effectiveness has proved by experiment.