A Novel Method for Small Object Tracking Based on Super-Resolution Convolutional Neural Network

Abstract

In this paper, we propose a novel tracking algorithm that can work robustly in small object tracking. The algorithm is inspired by recent advances in Super-Resolution Convolutional Neural Network (SRCNN). In contrast to most existing trackers which simply combine super-resolution (SR) and tracking, a deep network was proposed to better utilize SR features from different layers for small object tracking, which is rarely investigated in previous works. Specifically, we train a designed SRCNN offline to acquire SR features that are more robust against low resolution. Afterwards, this is followed by network weight transfer from offline training to the super-resolution tracker (SRT). The whole network can be tuned online to fit appearance variation. The SRT tracker employs a compact CNN structure without expanding layers or deconvolution layers and thus reduces the computational cost. Therefore, the tracking model achieves real-time performance while still maintaining good performance. The proposed method is evaluated on public tracking benchmarks, the simulation shows that SRT significantly increase the accuracy of tracking small objects.