Skeleton Based Person Location Tracking using Reinforcement Learning with an Expanded Action Space

College

College of Engineering & Technology (Batten)

Department

Electrical and Computer Engineering (ECE)

Graduate Level

Doctoral

Presentation Type

Poster Presentation

Abstract

Person location tracking is an important part of many security and surveillance applications. Typical tracking is accomplished using static visual sensors like security cameras to observe persons in a scene. However, a recent shift has been made away from strictly visual methods toward methods based on skeleton data. The use of skeleton data has already been greatly explored for many surveillance applications including action recognition and person re-identification. Skeleton data offers several benefits in this space including compact data representation, environmental noise tolerant, and resilience to long term scene change. While skeleton data remains popular for other many other human surveillance applications, few works have utilized skeleton data for person location tracking. Previous works for skeleton based person location tracking utilize deep reinforcement learning through the actor-critic design to create a skeleton based tracking policy. This policy is implemented using deep neural networks and offers several benefits, including, tolerance to noise that is typically present for skeleton data acquisition, and decreased reliance on large amounts of training data. However, the current implementation in this space utilizes an actor-critic approach with a limited, discrete action space, which restricts the operation of the algorithm. Furthermore, current works do not answer how actor-critic based approaches scale with action space size. In this work, we propose a study to understand how the actor-critic tracking approach responds with an expanded action space size. This expanded action space will enhance the network’s ability to make better and more refined decisions when tracking, resulting in a more refined tracking algorithm. This work provides insight into the actor-critic networks ability to scale to larger action spaces and serves as the basis for future research endeavors in this space.

Keywords

Skeleton Data, Deep Reinforcement Learning, Actor-Critic Design

This document is currently not available here.

Share

COinS
 

Skeleton Based Person Location Tracking using Reinforcement Learning with an Expanded Action Space

Person location tracking is an important part of many security and surveillance applications. Typical tracking is accomplished using static visual sensors like security cameras to observe persons in a scene. However, a recent shift has been made away from strictly visual methods toward methods based on skeleton data. The use of skeleton data has already been greatly explored for many surveillance applications including action recognition and person re-identification. Skeleton data offers several benefits in this space including compact data representation, environmental noise tolerant, and resilience to long term scene change. While skeleton data remains popular for other many other human surveillance applications, few works have utilized skeleton data for person location tracking. Previous works for skeleton based person location tracking utilize deep reinforcement learning through the actor-critic design to create a skeleton based tracking policy. This policy is implemented using deep neural networks and offers several benefits, including, tolerance to noise that is typically present for skeleton data acquisition, and decreased reliance on large amounts of training data. However, the current implementation in this space utilizes an actor-critic approach with a limited, discrete action space, which restricts the operation of the algorithm. Furthermore, current works do not answer how actor-critic based approaches scale with action space size. In this work, we propose a study to understand how the actor-critic tracking approach responds with an expanded action space size. This expanded action space will enhance the network’s ability to make better and more refined decisions when tracking, resulting in a more refined tracking algorithm. This work provides insight into the actor-critic networks ability to scale to larger action spaces and serves as the basis for future research endeavors in this space.