Deep learning assisted robust visual tracking with adaptive particle filtering

We propose a novel visual tracking algorithm based on the representations from a pre-trained Convolutional Neural Network (CNN). Our algorithm pre-trains a simplified CNN using a large set of videos with tracking ground truths to obtain a generic target representation. When tracking, Particle Filtering (PF) is combined to the fully-connected layer in the pre-trained CNN. Deep representations and hand-crafted features help to model tracking. To optimize the particles distribution, the velocity and acceleration information aids to calculate dynamic model. Meanwhile, our algorithm updates the tracking model in a lazy manner to avoid shift and expensive computation. As compared to previous methods, our results demonstrate superior performances in existing tracking benchmarks. A simplified deep learning frame is applied into visual tracking.Simple network frame and adaptive particle filter make the tracker more robust especially when the quick movement occurs.This method is validated through subjective and objective metrics.Quantitative comparisons highlight the contributions of FC features and hand-crafted features.