Cognitively Aligned Dual-Task Interaction for Small Object Perception in Highway Surveillance

Abstract

Small object perception requires tightly coupled semantic understanding and spatial reasoning. However, existing YOLOv8-based frameworks adopt decoupled heads, limiting task-level representation interaction and causing suboptimal cognitive consistency between classification and localization, particularly for low-resolution small objects. This paper proposes a cognitively aligned architecture with a novel Unified Task Interaction Head (UTI-Head) to enable structured interaction between classification and regression for joint semantic-spatial representation learning. It integrates cross-scale feature enhancement, task-adaptive decoupling, and dynamic-deformable geometric refinement to improve representation quality and task coordination. Experiments on the custom HSODD and public Mini-COCO datasets validate the method. Notably, the UTI-Head reduces parameters by nearly 20% compared with the YOLOv8 baseline while sustaining 85 FPS, demonstrating improved representation alignment and generalization capability in complex visual environments.