An Autonomous Feature Detection Method for Slow-Moving Small Target on Sea Surface Based on Kernelized Contextual Bandit

Currently, the contextual bandit (CB) has been applied in the field of slow-moving small target detection on sea surface, to solve the performance decline problem of feature detection method under less coherent pulse number. However, the existing method only gives a brief description on its implementation details and also has two shortcomings: 1) the algorithms used for estimating target location (TL) lack theoretical support and 2) the adopted policy is idealized. For the first issue, we assess the rationality and robustness of the initial TL estimation method, the online update algorithm for estimated TL, and the suppression algorithm for successive false alarms, both theoretically and experimentally. Then, we summarize the above three parts and propose a real-time TL estimation method. For the second issue, by modeling a nonlinear relationship between the expected reward of action and its observed context, the kernelized upper confidence bound (KernelUCB) policy is introduced. Subsequently, a k-nearest neighbor (KNN)-based approach is proposed to reduce its algorithm complexity. Furthermore, we propose designing an alternative action, which is enabled once main action chosen by the policy is invalid, thereby improving the performance of policy. By combining the above improvements with the CB process modeled from the classical feature detection process, the proposed method is obtained, which can realize the autonomous optimized selection for alternative feature detectors on each frame. The sufficient experiments demonstrate that the proposed method can integrate the performance advantages of various feature detectors while meeting the real-time detection requirement.