Abstract:
Majority of computer vision tasks, such as Human Activity Recognition (HAR), are
closely tied to security, virtual reality, video surveillance, and home monitoring
applications. This sets a new trend and turning point in the HAR system development
cycle. The problem is that most applications based on HAR available are low in
accuracy due to which the percentage of faulty recognition of human activities and false
alerts is quite high.
For the development of the system, we have selected UCF-SPORTS dataset
which contains 9118 images. Dataset consists of images having different activities,
which are classify according to the requirement of the system as their names. This
dataset is used to train the model. Multiple Transfer Learning models are used for the
development of this system. Design of the model is based on Xception, Inception, and
MobileNet along with stack of different layers to improve the accuracy. Method of
hyper-parameter tuning is used to analyse the behaviour of the discussed model. Model
is tested and trained for different range of values of Batch size, learning rate and epochs.
Performance of the model is validated through 5-fold cross validation. Experiments and
analysis show that model detects activities with almost 97% cross validation accuracy.
Human Activity Recognition System (HARS) is developed as a mobile
application for ease of use. In this system trained model is imported at the backend of
the application which is used to detect human activities. System detects activities by
vi
extracting frames from live feeds through camera. Moreover, system also notifies the
app user by alerts if there’s any suspicious activity to take safety measures as early as
possible. In this age where computers can now perceive and analyse their environment
with high accuracy, the detection system use latest and optimized model which
specializes in activity detection. . Application will be able to recognize and classify
several group of activities. This application is developed using Kivy a python
framework and ML/AI kit
