A High-Precision and Low-Latency Method for Pose Estimation Based on SAR-Net

Aiming at the challenges of poor real-time performance, insufficient occlusion robustness and limited model generalisation capability of 6D position estimation of target objects in complex scenes, this study proposes a multi-dimensional optimisation method, which aims to improve the accuracy, realtime performance and environment adaptability of position estimation. The main work includes the following aspects: (1) Propose a network architecture based on stand-in convolution with compression-excitation attention mechanism. The floatingpoint computation is reduced by the stand-in convolution, and the feature channel weights are dynamically adjusted in combination with the compression-excitation module, which enhances the model's ability to generalise to unseen objects. (2) For the occlusion problem, dynamic multi-directional convolution is designed to enhance the direction-awareness ability by diagonal direction convolution kernel. (3) The CATFusion module is proposed to reduce the computational complexity by lightweight feature fusion with cross-slice attention sparsification strategy.