Study of robust visual tracking based on traditional denoising methods and CNN

Visual object tracking has been widely used in many computer vision fields, ranging from intelligent video surveillance, autonomous driving, robot navigation, and human-computer interaction. However, complex external environmental interference may cause noise to the video images, during the entire process of data collection, transmission, and reception. In this paper, we aim to combine traditional denoising methods with the state-of-the-art tracker to implement robust tracking in noisy environment. In the first part, we apply salt-and-pepper noise and evaluate the performance of commonly used classical denoising algorithms. In the second part, we combine the denoising methods with Siamese Region Proposal Network to form a new tracker. The precision and success plots of different trackers are presented on noisy OTB100. Our study find that strong noise seriously affects the effect of target tracking and with use of denoising algorithm, the robustness of the tracker is enhanced greatly. In our experiment, for salt-and-pepper noise, median filter has obvious effect and can improve target tracking accuracy to the greatest extent.