Abstract:
Human Activity Recognition (HAR) refers to the detection and classification of different
types of human activities on the basis of data collected by sensors. Various sensory
devices are used for fetching the data such as accelerometers, microphones, gyroscopes,
and wearable devices to interpret and analyze human activities. The main goal of HAR is
to develop models and algorithms which can detect basic and complex human activities.
This paper presents the WISDM (Wireless Sensor Data Mining) 2019 data set collected
from 51 subjects performing 18 activities for a period of 3 minutes. The sensors used to
collect this data are accelerometers and gyroscopes and the devices used are smartphones
and smartwatches. For our thesis, we have developed a model similar to InceptionResNet
which can be used for different data sets to detect human activities accurately and
efficiently. The proposed model is an enhancement of an existing Inception ResNet
model and is more generic in nature as compared to the previous model. The model has
been tested with two input features 1D and 3D and 3 different batch sizes i.e.75, 100 and
200, which are the magnitude of 3D Accelerations, i.e., magw
a
’,and ’3D Accelerations,
i.e. (a
w,x a
w,y a
w)z’ . The previous model was developed and only tested on Smartphone
data of the WISDM 2011 dataset. To make our model more generic we have trained
and validated it on 3 versions of the WISDM 2019 dataset i.e. Smartphones data only,
Smartwatches data only, and, Smartphones and Smartwatches combined dataset. In our
proposed model, we have changed the architecture of the existing model by replacing
the GRU layer with the LSTM layer, reducing the number of inception models from 2
to 1, increasing the number of filters from 10 to 32,64 and, adding Max pooling layer
to the model. Time periods are set to 200 and step distance is set to 16. After making
these modifications to the model the accuracy achieved for all 18 activities in WISDM
2019 is from 97 % to 99.5 % using the sensors (gyroscope and accelerometer).
