Towards tiny object tracking for low-illumination wide-field

Abstract

Night is an inevitable scene for surveillance video. Due to the high image resolution, complex background, uneven illumination, and similarity between the target and the background of hawk-eye surveillance video, it is difficult for previous trackers to apply the tracking of a tiny object in such scenes. In this regard, this paper proposes to combine an online automatically and adaptively learning spatio-temporal regularized tracking algorithm with an efficient and effective low-light image enhancement algorithm to improve tracker performance. We constructed a new benchmark that includes 41 night surveillance sequences captured by Hawk-Eye cameras at night. Exhausted experiments have been conducted on this dataset, and the results show that by combining the two methods, the original algorithm can obtain better results in this dataset, and can meet the real-time object tracking, which contributes to the application of tiny object tracking in eagle-eye surveillance video at night.