Robust Generative Visual Tracker Based on Lion Swarm Optimization Algorithm

ABSTRACT Complex interferences such as target rotation, occlusion, and deformation increase the difficulty of the visual tracking process. Meanwhile, complex theoretical models and large tracking training networks increase computational power, leading to excessive reliance on expensive and complex hardware devices for real time tracking. To solve the above problems, this study develops a generative tracker via the lion swarm optimization algorithm (LSO for short). This method optimizes the population initialization process of LSO and the search strategy of lion groups through Circle chaotic mapping, thereby reducing the probability of being trapped in local extremes. In addition, this research designed adaptive adjustment factors for young lions based on the degree of iteration convergence and reinforcement learning based mother lion position compilation strategies, enabling the algorithm to automatically strengthen the global optimization in the early iteration and the local optimization in the late iteration, thereby improving the accuracy and efficiency of the LSO tracker in complex tracking interference scenes. Moreover, this research designed a frame size adjustment model to suppress the influence of background pixels and noise. Finally, the proposed method was qualitatively, quantitatively, and statistically compared with various classic trackers with evaluation benchmarks such as OTB2015 and VOT2018T as well as large‐scale benchmarks TrackingNet and LaSOT.