Target tracking by improved ECO

Abstract

ECO based trackers have achieved excellent performance on visual object tracking. However, Illumination variation and other factors still are challenging research problems in the process of tracking. Moreover, traditional neural networks also face information loss during the transmission process. In this paper, we introduce a new feature fusion (HE, FHOG-Encoder) and update strategy of learning rate. We propose an encoder network to extract features, which consists of two convolutional layers and three residual units. In addition, we design an updating strategy of learning rate, by computing absolute difference of inter-frame pixel, to effectively update sample space model. Experiments on challenging benchmarks OTB-100 are carried out. Experimental results show that our tracker achieves superior performance in some special cases, compared with the original ECO tracker.