abstract
- © 2019 IEEE.Action recognition has been highlighted by its implications on issues of security and integrity. However, although the dense trajectory method has achieved outstanding results, its implementation is associated with a high computational cost, creating the need for faster methods if we want to use these methods to protect our safety and integrity in real-Time systems.In this work, we explore the subject keypoints estimation to create what we call action recognition by key trajectories, which is a faster version of the dense trajectories approach. We tested our proposal on the KTH and UCF11 datasets and got as a result that our proposal recognizes subject actions in videos eight times faster with comparable results to the dense trajectories approach. i.e., 94.20% for the Key Trajectories compared to 95.65% of the dense trajectories in the KTH dataset and 80.66% for the Key Trajectories compared to 84.08% of the dense trajectories in the UCF11.