Relative Object Tracking Algorithm Based on Convolutional Neural Network for Visible and Infrared Video Sequences

In this paper, a novel relative object tracking algorithm using a convolutional neural network is proposed aiming to boost the tracking performance. A two-layer convolutional neural network extracts sparse feature representation of visible and infrared sequences via convolutional filters. The convolutional filters contain two types, object filter, and relative filters. In the first frame, we employ a set of normalized fusion patches as the object filters. Moreover, a relative model is explored to generate relative filters using k-means algorithms, which integrates information from both foreground and background to build accurate appearance model. This algorithm without training is robust and efficient. Quantitative and qualitative evaluations demonstrate that the performance of this algorithm improves significantly over the state-of-the-art techniques when applied to public testing sequences.