Real-Time Online Tracking via a Convolution-Based Complementary Model

Abstract

Several object tracking convolution networks have been proposed in recent years. Despite their favorable performance, the balancing of tracking accuracy and efficiency remains challenging. In this paper, we propose a real-time online tracking method based on complementary tracking models: the convolution-based discriminative model (CDM) that can predict the center location of an object and the convolution-based generative model (CGM) that estimates the scale of the target. In the CDM model, we leverage a simple convolution operation to model the correlation between the apparent features (gradient and color features) of the object and its background. Then, the center location is predicted by maximizing the response value of the convolution. In the CGM model, a two-layer convolution network is proposed to learn geometric structural information, and the target scale is estimated by selecting the best candidate extracted from the foreground of the target through the observation model. Moreover, online updating and the fast Fourier transform are adopted for fast learning and detection. Despite its surprisingly lightweight structure, the proposed tracker performs favorably against the state-of-the-art methods in terms of efficiency, accuracy, and robustness on the CVPR2013 tracking benchmark data set.