<p>Sound events often occur in polyphonic form. Multiple overlapped events with similar frequencies may pose a significant challenge for sound event detection to distinguish the boundaries of events and identify their categories. Conventional time-frequency representations struggle to capture fine-grained temporal dynamics and spectral diversity under such conditions. To address this problem, we propose a sound event detection method that integrates frame-level manual features with acoustic features. The manual feature module extracts seven features that explicitly target temporal boundary ambiguity, spectral overlap, and signal complexity in polyphonic scenes. To more effectively fuse the two types of features, we introduce a cross-attention mechanism to achieve precise alignment between manual and acoustic features. The fused representation is subsequently processed by an Omni-dimensional Dynamic Convolution Network, which adaptively extracts discriminative patterns across multiple dimensions. Experimental results on the DESED show that the proposed method outperforms baseline systems, confirming its effectiveness in sound event detection tasks.</p>