Role Identification from Human Activity Videos using Recurrent Neural Networks

Anam Arshad, Vivek Tiwari, Mayank Lovanshi, Rahul Shrivastava

Abstract:

Recognizing the underlying roles in mutual activity
is more informative. Role identification h as t he p otential to
improve wide range of applications, of activity recognition from
safety and security to healthcare. In recent research, for role
identification, w ork i s d one t o i dentify r oles b y c apturing the
knowledge of body parts from an image. This work is complex
and not sufficient to take input as English sentences and capture
the sequencing and relationship between words. There is a
need for simple work which could use recent technologies like
Recurrent Neural Networks to capture the recurrent nature of
sentences to identify roles. The contribution of this work is
proposing a Computational Long Term Memory model where
sentences are stored as features and given to a Recurrent Neural
Network to identify roles. The appropriate dataset is not available
for role identification u sing s entences. I n t his v iew, t his work
attempt to develop a new custom dataset. The proposed model is
tested on accuracy using various Recurrent Neural Networks like
Recurrent Neural Network (RNN), Long Short Term Memory
(LSTM), Gated Recurrent Units (GRU) etc. The LSTM model
gave effective accuracy of 60% on the small custom dataset.
Index Terms—Role identification, Recurrent Neural Networks,
Long Short Term Memory, Gated Recurrent Units.