Pasiam: Predicting Attention Inspired Siamese Network, For Space-Borne Satellite Video Tracking

Jia Shao,Bo Du,Chen Wu,Pingkun Yan

DOI: https://doi.org/10.1109/ICME.2019.00260

2019-01-01

Abstract:Tracking a moving target of interests from a space-borne satellite video is really challenging. The difficulty lies in that the target usually occupies only several pixels, so that its features are very difficult to obtain. Besides, appearance features of the target would be unobvious when it is occluded, suffers from illumination variation influence or moves to similarity surroundings. In this paper, we propose a PREDICTING ATTENTION Inspired SIAMESE NETWORK (PASiam) for space-borne satellite video tracking, which constructs a fully convolutional Siamese network with shallow-layer features to obtain fine-grained appearance features. Moreover, a predicting attention is proposed to deal with occlusion and obscure. It employs Gaussian mixture models (GMM) to detect the target's motion status, and Kalman filter to predict and correct the target's location. Quantitative evaluations are performed on three real satellite video datasets. The results show our approach outperforms the state-of-the-art tracking methods while running at 54.83 FPS.