Aircraft tracking based on fully conventional network and Kalman filter

Abstract

Aircraft tracking is a significant technology for military reconnaissance, but there is no efficient algorithm to solve this particular problem. Recently, research based on deep learning for object tracking has developed rapidly, and the performance is greatly improved compared to the traditional methods, so the authors refer to relevant work and make an improvement on the previous research to improve the performance on aircraft tracking. They first learn the idea from region-based fully convolutional networks to perform detection on each frame of video. To avoid the target drift due to the failure of object detection on a certain frame, then they employ Kalman filter (KF) and extended KF together to predict the moving trajectory of the target. Beyond that, this method can confine the valid range based on the size of a target object, which increases the speed of detection. This approach can also correct the bounding box on adjacent frames. The steps are not complicated but have an excellent performance. Through the experiment, it is clear that the proposed method is reasonable and more precise.