Visual Object Tracking Based On Convolutional Neural Networks

Visual object tracking is a hot topic in computer vision research in recent years.Tracking algorithms are expected to robustly localize the object of interest in videos,for supporting various applications such as self-driving cars.Deep learning was widely used in computer vision,trackers which sort the power of convolutional neural networks outperform classical ones on tracking benchmarks.Meanwhile,these deep trackers are also bigger and hard to run in real-time.We surveyed state-of-art algorithms to find the way to build the light-weight,fast,stable single object visual tracker.In this paper,we proposed a single object visual tracking algorithm based on fully convolutional neural networks,correlation filtering and regional search strategy,the tracker is comparatively light-weight and run in real-time.Generally,the paper covers three aspects of our work.1)A specifically designed model Net-A and corresponding loss function were employed to be trained on Image Net VID dataset and the algorithm outperforms several algorithms proposed in recent years.2)In this paper,we proved that AUC equals to AO theoretically,and all the analysis work was based on algorithms' AO score.We also proposed a fusing methodology of tracking algorithms based on the fully convolutional neural network.3)Inspired by the analysis results,we proposed to use a pre-trained model trained for image classification task,freeze its weights,adding new layers to form Net-B,after trained in the same way as Net-A,the two networks would serve in a two-stream architecture for visual object tracking.Experimental results showed that our algorithm outperforms Siamese FC and GOTURN which are also real-time trackers.When comparing to slower trackers,it is only inferior to MDNet while it may not be trapped by some over-sampling issues due to its simplicity which is also a virtue.