Mamba-AANet: Mamba-Based Temporal Modeling for High-Precision Arm Pose Tracking

Target pose tracking is a fundamental task in computer vision, yet it remains challenging due to issues such as sudden illumination variations and occlusions. Current methods often face a trade-off between accuracy and computational efficiency.Traditional geometry-based methods lack adaptability, while Transformer-based approaches have quadratic complexity issues. To address these limitations, this paper presents Mamba-AANet, a novel framework that incorporates the Mamba state-space model to achieve efficient long-term temporal modeling with linear complexity. A multi-scale adaptive mask mechanism is designed to dynamically adjust the receptive field at various temporal scales. Furthermore, the model includes a feature enhancement module that integrates bidirectional information flow to capture comprehensive contextual dependencies. This module is also equipped with a Memory Enhanced Mechanism for maintaining long-range interactions, cross-layer feature enhancements to enrich representation learning, and a Quality Aware Dynamic Fusion mechanism for adaptive feature fusion based on perceptual quality. Experiments on datasets such as Panda Syn, Panda 3CAM - RS/AK, and Panda Orb show it outperforms methods like CenterNet and CenterTrack in metrics like PCK and median@pix, with ablation experiments verifying the effectiveness of key components, providing a new paradigm for robotic pose tracking in dynamic scenarios.