Article ID: 2024EAL2088
To address the challenges of low detection accuracy resulting from occlusion and scale variation in complex traffic scenarios, as well as the high computational complexity and large model parameters associated with traditional methods, this paper proposes a Lightweight Cross-Scale Feature Fusion Algorithm. Firstly, we design the Lightweight Cross-Scale Feature Fusion Module (LCFM), which incorporates an improved internal fusion block to facilitate interactive feature fusion. This design enhances the model's adaptability to occlusion and scale change while reducing the number of input feature channels to make the model more lightweight. Furthermore, by integrating Squeeze-and-Excitation (SE) attention with multi-branch convolution operations from the Inception structure, the model can more accurately capture multi-scale object features. Additionally, Linear Deformable Convolution (LDConv) is employed to adaptively handle shape changes through offset learning, thereby reducing computational redundancy and improving the model's overall adaptability.
