Object Detection and Tracking based on Recurrent Neural Networks

Abstract

Robust and accurate visual tracking is one of the most challenging computer vision problems nowadays. The development of deep learning improves the performance of tracking, but in non-constrained scenarios, the accuracy will decrease dramatically. To tackle this issue, we put forward our own tracking framework on the basis of recurrent neural networks (RNN). Firstly, we use the temporal correlation of RNN to predict the direction of the object at next frame. Secondly, we determine region of interest (ROI) using the direction predicted by RNN. And then, we detect the object within ROI. Finally, the output of the object detection model is transmitted to the direction prediction model, which is the input of the next frame, for realizing object tracking framework. We perform comprehensive experiments on two benchmark datasets: OTB and VOT2016. Our approach achieves an accuracy of 0.826 and a speed of 41fps on OTB100, which is effective and efficient.