Efficient Recurrent Detection for Visual Tracking

In this paper, we propose a visual tracking algorithm, which is called the recurrent detection using two-stage clip training (RDTCT) algorithm, based on the detection model pretrained on large datasets and convolutional recurrent neural network (ConvRNN) cells. The proposed model is composed of two branches. The initial state extraction branch is to extract the initial state of ConvRNN cells and the prediction branch is to predict the bounding-boxes based on the current image and previous states. Since the advanced network architecture is applied, the proposed algorithm has extremely high computation efficiency. When training the proposed model, conventional back-propagation through time does not work well due to inefficient utilization of video data. Thus, a two stage clip training is proposed to replace back-propagation through time. Experiments show that the proposed tracker can handle multiple circumstances and reduce the time complexity to around 25 frames per second (fps).