A Lightweight and Real-Time Network for Unmanned Aerial Vehicle Object Tracking

Unmanned Aerial Vehicle (UAV) object tracking presents a promising application scenario for both military and civilian domains. However, the computational demands of state-of-the-art trackers present a formidable obstacle to achieving real-time operations on embedded platforms in small UAVs. Although some lightweight backbone networks, such as Mobile-Net and Shuffle-Net, have been used in the Siamese network, they still cannot achieve real-time tracking. Therefore, we adopt a more lightweight feature extraction network and prune both the backbone network and head sub-network to ensure the real-time operation of the tracker on embedded platforms. In contrast to the Mobile-Net used in classification and detection networks, we employ a large number of convolution kernels with larger receptive fields and multi-level feature fusion output to address challenges posed by small target size and scale changes in UAV tracking tasks. Additionally, The pixel-wise correlation is particularly effective in enhancing the representational capacity of feature correlation in scenes with out-of-plane rotation, making it better suited for UAV target tracking tasks. The qualitative and quantitative experimental results demonstrate the effectiveness of our tracker in enhancing UAV tracking performance, making it a suitable option for deployment on small UAVs.