Multidimensional Feature Enhancement and Interactive Fusion Method for Collaborative Perception

Focused on the issue of failing to meet the requirement of precise perception and low model complexity in existing intermediate collaborative perception methods, this study proposes a novel multidimensional feature enhancement and interactive fusion method for collaborative perception. First, the method enhances features related to object through feature enhancement and introduces a Multi-dimensional Self-interaction Enhancement Module (MSEM), which performs parallel and linear hybrid interactions across spatial and channel dimensions to effectively amplify the prominence of target features. Next, multiscale features are extracted with different downsampling rates. A Cross-agent Multidimensional Interaction Fusion Module (CMIFM) is then employed to promote inter-agent feature interaction and fusion among different agents in both spatial and channel dimensions. Experiments on the large-scale DAIR-V2X-C and V2XSET datasets show that, compared to the baseline Pyramid algorithm (which has extremely low parameters), the proposed method achieves significant performance gains with only a 0.37% increase in parameters. Specifically, on DAIR-V2X-C, it improves $\mathrm{AP} \text{@} 0.5$ and $\mathrm{AP} \text{@} 0.7$ by 0.96% and 3.21%, respectively; and on V2XSET, it enhances AP@ 0.5 and AP@ 0.7 by 2.02% and 1.51%, respectively. Thus, this method can remarkably boost perception accuracy while maintaining a very low number of parameters.