TrackingMamba: Visual State Space Model for Object Tracking
Qingwang Wang, Liyao Zhou, Pengcheng Jin, Xin Qu, Hangwei Zhong, Haochen Song, Tao Shen
Abstract
In recent years, UAV object tracking has provided technical support across various fields. Most existing work relies on convolutional neural networks (CNNs) or visual transformers. However, CNNs have limited receptive fields, resulting in suboptimal performance, while transformers require substantial computational resources, making training and inference challenging. Mountainous and jungle environments-critical components of the Earth's surface and key scenarios for UAV object tracking-present unique challenges due to steep terrain, dense vegetation, and rapidly changing weather conditions, which complicate UAV tracking. The lack of relevant datasets further reduces tracking accuracy. This article introduces a new tracking framework based on a state-space model called TrackingMamba, which uses a single-stream tracking architecture with Vision Mamba as its backbone. TrackingMamba not only matches transformer-based trackers in global feature extraction and long-range dependence modeling but also maintains computational efficiency with linear growth. Compared to other advanced trackers, TrackingMamba delivers higher accuracy with a simpler model framework, fewer parameters, and reduced FLOPs. Specifically, on the UAV123 benchmark, TrackingMamba outperforms the baseline model OSTtrack-256, improving AUC by 2.59% and Precision by 4.42%, while reducing parameters by 95.52% and FLOPs by 95.02%. The article also evaluates the performance and shortcomings of TrackingMamba and other advanced trackers in the complex and critical context of jungle environments, and it explores potential future research directions in UAV jungle object tracking.