Lightweight siamese object tracking algorithm based on SiamBAN

Abstract

The UAV platform has limited computing resources, and the tracking algorithm needs better speed and accuracy tradeoff, a lightweight siamese network target tracking algorithm called SiamBAN-T based on SiamBAN. Firstly, to reduce the number of network parameters, mobilenetV3 was as to extract the siamese feature. Secondly, we introduce CA attention into the feature fusion module to enhance perception ability regarding target spatial-position information. Thirdly, multibranch cross correlation is incorporated into the head of the network to strengthen boundary information and scale information, thereby improving the anti-interference capability of our trackier. Finally, a feature enhancement module is designed to improve classification and regression abilities. Experimental results on UAV123 dataset demonstrate that compared with the original algorithm, our improved algorithm achieves an increase in success rate by 0.8% and accuracy by 0.8%. The running speed has been enhanced by 7.6 times for PC devices and 18.5 times for airborne mobile terminals, respectively. These experimental findings indicate that our SiamBAN-T significantly enhances tracking speed while maintaining high precision.