.png){kind=icon}
.png){kind=icon}
Return
AFMtrack: Attention-Based Feature Matching for Multiple Object Tracking
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2024.3411617.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Real-time multiple object tracking plays a pivotal role in autonomous driving applications, particularly in real-world applications. Current methods in various domains often face an inherent trade-off between accuracy and speed. This dilemma arises from the need to achieve high precision, which tends to entail the development of more intricate models. However, these complex models often incur a high processing overhead. In this study, we introduced an Attention-based Feature Matching tracker for multiple-object tracking, named AFMTrack. We developed AFMTrack as a Transformer-based tracking method that improves speed and accuracy to provide a comprehensive solution to real-time tracking challenges. In particular, we designed a feature-matching module, which is a multi-layer attention-based network, to produce an association matrix by learning the correspondence between frames. Furthermore, in simple scenarios, we realized that a few layers were sufficient to attain the desired accuracy. Therefore, an Early stopping mechanism is added to halt the process when early layers produce confident predictions. This significantly accelerated AFMTrack without compromising accuracy. The effectiveness of AFMTrack is proven with 89.64% MOTA, 132 FPS on the KITTI dataset, and 79.3% MOTA, 31 FPS on MOT17 dataset, surpassing most of methods on the leaderboards. The evaluations provide solid evidence that AFMTrack excels in multiple-object tracking, achieving state-of-the-art performance in this domain.
Keywords:
Transformers
Feature extraction
Target tracking
Real-time systems
Task analysis
Object detection
Detectors
Autonomous driving
Object tracking
multiple-object tracking
transformer
self-attention. cross-attention