A location-aware siamese network for high-speed visual tracking

Accurately locating the target position is a challenging task during high-speed visual tracking. Most Siamese trackers based on shallow networks can maintain a fast speed, but they have poor positioning performance. The underlying reason for this is that the appearance features extracted from the shallow network are not effective enough, making it difficult to accurately locate the target from the complex background. Therefore, we present a location-aware Siamese network to address this issue. Specifically, we propose a novel context enhancement module (CEM), which contributes to capturing distinguished object information from both the local and the global levels. At the local level, the features of image local blocks contain more discriminative information that is conductive to locating the target. At the global level, global context information can effectively model long-range dependency, meaning that our tracker can better understand the tracking scene. Then, we construct a well-designed feature fusion network (F-net) to make full use of context information at different scales, where the location block can dynamically adjust to the convolution direction according to the geometry of the target. Finally, Distance-IoU loss (DIoU) is employed to guide the tracker to obtain a more accurate estimation of the target position. Extensive experiments on seven benchmarks including the VOT2016, VOT2018, VOT2019, OTB50, OTB100, UAV123 and LaSOT demonstrate that our tracker achieves competitive results while running at over 200 frames-per-second (FPS).