To address the dynamic occlusion problem caused by stalk tilting in automated corn harvesting scenarios, this paper proposes a lightweight corn ear detection algorithm named OASA-YOLO. By integrating an Occlusion-Aware Multi-Scale Module (OAMS-Block) and an Adaptive Heterogeneous Kernel Selection Protocol (Adaptive-HKS), the algorithm significantly enhances detection robustness in complex occlusion environments. The OAMS-Block employs Stochastic Occlusion-Aware Convolution (SOA) and Global Query Learning (GQL) mechanisms to dynamically suppress occlusion noise and strengthen multi-scale feature propagation. Based on local occlusion density map, adaptive HKS adaptively selects convolution kernel to balance local details and global semantics. Experiments demonstrate that OASA-YOLO achieves a mean average precision (mAP) of 96.3% on a self-built corn ear dataset, outperforming YOLOv9 (90.7%) and YOLOv11 (91.8%) by 5.6% and 4.5%, respectively. With only 2.6M parameters and 3.9G FLOPs (floating point operations), the model meets the real-time requirements of embedded devices.