Deep Visual Detection-Tracker for Aerial Target
Wei Mei, Xiaoyu Feng, Ruifei Yu, Dashuai Hu · 2019
Visual tracking of aerial target in complex battle environments is a challenging task. To meet this challenge, we develop a deep visual detection-tracker, which incorporates a deep learning detector, a correlator and a Kalman filter. The powerful capability of existing visual tracking algorithm based on deep learning attributes to a robust discriminative appearance model, which is established by integrating hierarchical convolutional features from a deep convolutional network. For our task a series of measures are applied to insure the tracking performance. First, the performance of the detector, in terms of both accuracy and speed, is improved by using techniques of local labeling for training dataset, adaptive threshold for detection scores and quantity compression of region proposals. Second, to eliminate the distracters, a brand-new correlator of target attributes and pooling features is designed for calculating matching degrees of the target and candidates. Besides, a Kalman filter is used to produce a reasonable prediction when the detector fails to detect the target. To demonstrate our deep tracker, a customized videos /images dataset of aerial target is built for training and test. Experiment results show that the developed deep visual detection-tracker is more robust and computationally more efficient than existing tracking algorithms based on deep learning.