A Strong UAV Vision Tracker Based on Deep Broad Learning System and Correlation Filter

Object detection and tracking is always a challenging issue in UAV (unmanned aerial vehicle) application. Especially, in the scene of UAV-ASV (autonomous surface vehicle) cooperative system, UAV vision based target tracking performance has been suffering from the target rotation and fast motion. Aiming at optimizing the related UAV vision tracking performance, a SDSST tracker(strong discriminative scale space tracking) with automatic initialization and self-adjusting is developed in this paper. Firstly, in the step of initialization, combining the advantages of both fast optimization of BLS (Broad Learning System) and efficient image processing of CNN (convolutional neural networks), a novel DBLS (Deep Broad Learning System) is posed for the target detection. Meanwhile, a Q-learning based DBLS architecture searching is further introduced. Then, in terms of width/height ratio self-adjusting, this article proposed a novel filter state supervisor that helps to find the abnormal estimated state caused by rotation in target scale estimation. Basically, this so called filter state supervisor could take RSV (Rolling Standard Value) as input feature and give out the filter state. Finally, the abnormal filter state would be adjusted by an appropriate alternative by searching in the proposed rotation angles memory, so that an optimized self-adjusting could be realized. Meanwhile, extensive experiments are performed on data set of USV center in Qiandao Lake, yielding a competitive result compared with five other prevalent trackers. Note to Practitioners—This paper was motivated by the problem of target lost caused by sudden change of target motion in the UAV-ASV vision tracking. Usually, the rotation motion of ASV is a major operation when ASV carries out a maritime assignment. However, the poor tracking state supervision and inflexible scale updating method as well as manual initialization in the existing approaches lead to inappropriate scale and target lost when the target undergoes rotational motion. Therefor, this paper proposes a strong vision tracker (SDSST) to enhance the original DSST in the three aspects: automatic initialization, filter state supervisor, and self-adjusting. This can allow the tracker to initialize without human interference, also bad tracker state can be informed by filter state supervisor. When bad state happens that the target scale in the tracker will be updated flexibly based on proposed rotation angles memory. Finally, The proposed method is implemented on the real filed UAV vision data collected by UAV-ASV system in the Qiandao Lake. The results show that SDSST achieves competitive result compared to 7 other prevalent trackers.