Visual Tracking via Hierarchical Deep Reinforcement Learning

Visual tracking has achieved great progress due to numerous different algorithms. However, deep trackers based on classification or Siamese network still have their specific limitations. In this work, we show how teach machines track a generic object in videos like humans, who can use few search steps perform tracking. By constructing Markov decision process Deep Reinforcement Learning (DRL), our agents learn determine hierarchical decisions mode and motion estimation. To be specific, Hierarchical DRL framework is composed of Siamese-based observation which models the information an arbitrary target, policy for switch actor-critic box regression. This strategy more line with human behavior paradigm, effective efficient cope fast motion, background clutter large deformations. Extensive experiments GOT-10k, OTB-100, UAV-123, VOT LaSOT benchmarks, demonstrate that proposed tracker achieves state-of-the-art performance while running real-time.