Data Association With Camera Parameters Estimation for Object Tracking From Drones
paper Menu
This paper considers the problem of inaccurate measurement-to track association (M2TA) and poor tracking caused by camera motion changes in drone-captured video. The camera often changes its field of view to track targets; however, the sudden change leads to inaccurate M2TA and degrades tracking performance. Previous work estimated the 3D camera motion parameter vector (zoom ratio, panning, and tilting) and associated measurements and tracks only between two consecutive frames. This paper extends the camera motion parameter to 4D by including rolling and sequentially associates (forward) measurements to tracks over the entire data. The estimated camera parameters improve the predicted measurements and achieve better M2TA. Results on real data illustrate the benefits of the proposed method (association with 4D camera parameters estimation) that yields better associations and improves tracking accuracy compared to the state-of-the-art gating method based on inflated covariances.
