Selective Multi-Scale Feature Fusion Network for Object Detection in Autonomous Driving

Object detection is essential for autonomous driving, as it provides crucial perception of surrounding environments by locating and classifying semantic targets to support trajectory planning and ensure driving safety. However, traditional methods still exhibit insufficient boundary representation and ineffective multi-scale feature fusion, leading to suboptimal performance in complex scenes. To address these issues, we propose a selective multi-scale feature fusion network built upon a single-stage detection framework. The framework integrates three main components: a Local-Enhanced Global Modeling (LEGM) module that combines convolution and self-attention to strengthen multi-scale feature representation, a Selective Boundary Aggregation (SBA) module that enhances contour information and deep semantics, and a lightweight Transformer-based decoder that adaptively filters queries to improve robustness. This design ensures modular flexibility, training stability, and improved detection accuracy. Experiments on widely used benchmarks demonstrate that the proposed method achieves high localization accuracy while maintaining near real-time inference. Compared with YOLOv13, the mAP@0.5 improves by 4.8%, small-object detection accuracy increases by 12.8%, and performance improves by more than 5% under challenging conditions such as low illumination and occlusion.