PROJECT TITLE :

Hierarchical Tracking by Reinforcement Learning-Based Searching and Coarse-to-Fine Verifying

ABSTRACT:

A class-agnostic tracker typically has three main components, namely its motion model, its target appearance model, and its updating technique. Aside from these complex appearance models and updating methodologies, the most current top-performing trackers tend to focus on relatively simple motion models that may result in an inefficient search and worse tracking performance. Hierarchical tracking is proposed to solve this problem by using data-driven search at the coarse level and fine level verification to learn how to move and track. The coarse location of an object is provided by a data-driven motion model developed using deep recurrent Reinforcement Learning. As an action-decision issue in reinforcement learning, our tracker uses a deep Q-network based on a recurrent Convolutional Neural Network to learn effective data-driven searching rules for motion searches. In addition to reducing the search space, the motion model developed can also provide more trustworthy interest areas for additional verification. Kernelized correlation filter (KCF)-based appearance model is used to evaluate a local region centred on the projected position from the motion model at the fine level. Circulant matrices and rapid Fourier transforms allow the KCF-based appearance model to efficiently evaluate huge numbers of candidate samples in the local region. Finally, an estimator that is both simple and reliable is created to assess the likelihood of tracking failure. That our tracker performs better than current trackers may be demonstrated by the results of our tests on OTB50 and OTB100.