Feature semantic self-attention mechanism target detection tracking method for unmanned aerial vehicle aerial video

Abstract

The invention discloses a feature semantic self-attention mechanism target detection tracking method for an unmanned aerial vehicle aerial video. In a target detection stage, a feature extraction network based on hierarchical attention enhancement is provided, then, a feature fusion network based on multilayer context semantic enhancement is constructed, and a context guide module is introduced in each level of feature output layer; and finally, introducing a loss function of the central point distance into the loss function. In a target tracking stage, a lightweight visual self-attention re-recognition network is provided, and a lightweight visual Transform network is introduced; in addition, a tracking association matching method based on a distance intersection-union ratio is optimized. According to the method, the local and global perception capability and the multi-scale target detection performance of the target detection and tracking model are remarkably improved, and the sensitivity of the network to the dense arrangement of the targets is enhanced.

Images