PFSAN: An Efficient Perception Network for Autonomous Driving Based on Progressive Feature Splitting and Aggregation

Bottleneck structures are widely used in object detection networks but often face challenges related to limited feature representation and computational efficiency, especially in lightweight designs. To address this, this paper introduces a novel Progressive Feature Splitting and Aggregation Module (PFSAM) that improves feature extraction through feature decomposition and aggregation. It also integrates structural re-parameterization to reduce computational redundancy and optimize gradient flow. Using PFSAM, several network variants are designed to meet varying computational resource requirements. Experimental results demonstrate that the PFSAM module achieves the fastest inference speed and exceptional efficiency with minimal computational complexity and parameter count. In the BDD100K dataset, the PFSAN network excels in small object detection and complex background scenarios, striking a balance between lightweight design and high performance. It provides an efficient solution for object detection tasks.