CPD-KD: a cooperative perception network for discrepancy feature fusion through knowledge distillation

The environmental perception of intelligent connected vehicles is a new perception paradigm centered around roadside sensing devices and 5G wireless communication. This approach expands the perception range of onboard sensors from localized observation to a global view, significantly enhancing the accuracy of vehicle perception. Traditional perception algorithms often face issues in real-world datasets, such as blurriness of target features from the vehicle's perspective and loss of specific information when integrating vehicle and infrastructure features. To address these issues, we propose a Cooperative Perception network for Discrepancy feature fusion through Knowledge Distillation, named CPD-KD. The contributions of this work are threefold: First, we introduce a Sparse convolution-based Knowledge Distillation network (SKD) to alleviate blurring of single-view point cloud features by incorporating prior knowledge from fused viewpoint point cloud features. Second, we design a Discrepancy Feature Attention Fusion module to enhance Cooperative efficiency of differential information between vehicles and infrastructures. Finally, we validate the CPD-KD algorithm on the real-world dataset DAIR-V2X and the simulated dataset V2X-Set. Experimental results show that the proposed CPD-KD algorithm effectively enhances the accuracy of vehicular and infrastructural Cooperative perception.