Object tracking using simple recurrent neural network with fast image processing speed

Abstract

Real-time Object detection and real-time object tracking have taken a large part in the traditional field of computer vision. However, despite many advances in technology, these areas remain still challenging areas for reasons of object occlusion, blurring, and fast motions. Basically there are two methods of tracking objects: tracking-bydetection and template matching. The tracking-by-detection method is a method of sampling many bounding boxes around the target and learning how to overlap the target (IOU). The template matching method works by storing the target patch of the first or previous frame and then finding the best match with the target in the current frame. In this section, we are continuously studying the external shape change by using the simple recurrent unit (SRU), which is widely used in the natural language processing (NLP) area due to the recent light and fast learning speed, We will apply this to the tracking-by-detection algorithm, which has advantages over template matching in terms of object occlusion, object blurring, and fast motions. This method has better performance in real-time performance due to its light weight network characteristics with advantages in occlusion, blur, and fast motion compared to other conventional recurrent neural network methods and other template matching tracking methods in TB-30 dataset. We also confirm that IOU measurement has better or similar performance than tracking networks using other recurrent neural networks.