Small target unmanned aerial vehicle dynamic tracking method and system based on multi-scale attention

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) technology, and more particularly to a method and system for dynamic tracking of small targets by UAVs based on multi-scale attention. The method includes: extracting multi-scale features from the input image using a lightweight multi-scale backbone network; inputting the multi-scale features into a collaborative spatial-channel attention fusion module for feature enhancement and fusion to obtain enhanced fused features; updating the memory state and predicting the short-term state of the target through a dynamic memory and state prediction network based on the enhanced fused features and the historical tracking state; and outputting the target's bounding box and tracking confidence in the current frame in parallel using a localization and classification head based on the state prior from the dynamic memory and state prediction network and the enhanced fused features. This invention solves the technical problems of detail loss, difficulty in adapting to scale changes, poor robustness to dynamic interference, and the need for lightweight algorithms in small target tracking from the UAV's perspective.

Images