Online Multi-Object Tracking Using Selective Deep Appearance Matching
Young-Chul Yoon, Young-Min Song, Kwangjin Yoon, Moongu Jeon · 2018
In this paper, we focus on designing appearance matching network and solving computational bottleneck problem of it. From the development of deep neural network and graphic device(GPU), many research topics in computer vision (e.g. detection, classification) achieved state-of-the-art performance using convolutional neural network(CNN). In multi-object tracking, also, there have been several works which used CNN for extracting appearance feature of targets. Although, deep appearance feature improved an accuracy of tracking, it increased processing time and made an algorithm hard to be applied in real-world situation. So, we propose a simple technique to improve speed by removing redundant appearance matchings. Also, we propose a structure of joint-input siamese network and method to train it. We verify the performance of our work by comparison with recent online trackers.