Abstract
Visual analytics applications often rely on target
tracking across a network of cameras for inference
and prediction. A network of cameras generates
immense amount of video data and processing it for
tracking a target is highly computationally expensive. Related works typically use data association
and visual re-identification techniques to match target templates across multiple cameras. In this thesis, I propose to formulate this scheduling problem
as a Markov Decision Process (MDP) and present
a reinforcement learning based solution to schedule
cameras by selecting one where the target is most
likely to appear next. The proposed approach can
be learned directly from data and doesn’t require
any information of the camera network topology.
NLPR MCT and DukeMTMC datasets are used to
show that the proposed policy significantly reduces
the number of frames to be processed for tracking
and identifies the camera schedule with high accuracy as compared to the related approaches. Finally, I will be formulating an end-to-end pipeline
for target tracking that will learn a policy to find
the camera schedule and to track the target in the
individual camera frames of the schedule