Résumé: Siamese network-based trackers have been proven to maintain splendid performance. Recently, visual tracking has been applied in unmanned aerial vehicle(UAV) tasks. However, it is a challenging task because of the influences by aspect ratio changes, out-of-view and scale variation, etc. Some Siamese-based trackers ignore context-related information generated in the time dimension of continuous frames, lose a lot of foreground information and generate redundant background information. In this paper, we propose a novel the feature fusion network based on convolutional self-attention blocks. The convolutional self-attention blocks are composed of ResNet bottleneck blocks with multi-head self-attention (MHSA) blocks. We eliminate the spatial (3×3) convolution operator limitation through the MHSA blocks in the last stage bottleneck blocks of ResNet. Convolutional self-attention blocks capture the global context-related information of the given target images and further improve the accuracy of global match between a given target and a search region. Extensive experimental evaluations on OTB2015 and four UAV benchmarks, i.e., UAV123, UAV20L, DTB70 and UAV123@10fps. The experimental results demonstrate that the proposed tracker can achieve excellent performances against SOTA trackers for UAV tracking and lead to real-time average tracking speed of 181fps on a single GPU.

Langue: Anglais