Description
The evolution of an innovative technologies in the field of connected things, edge computing and predictive machine learning learning algorithms have made the world an easier place to reside for humans. Detection of human activities and gestures have been an active research topic in many industries which are trying to develop smart applications like smart healthcare, smart surveillance, smart home automation, etc. as the demand is quite high. In this thesis, we have implemented an efficient, inexpensive and effective approach to analyze and detect different activities done by multiple people in an environment. The work done in this thesis helps us estimate the correct number of people in an environment. It develops different methods to detect different human activities in real-time. For the data collection, we used multiple low-resolution thermal sensors, with 4x16 resolution each. The sensors were placed in a small room, with dimensions 2.5 meters x 2.75 meters, where there were two or three people in different times. To estimate the number of people, we took one sensor and performed experiments by placing it on different walls. The results showed that Connected Component method was a better algorithm when compared to Window Size method, with an efficiency (accurate prediction of number of people) of 100% when the sensor was placed on the z-axis wall. Furthermore, to detect activities performed by different people, we collected data when there were either two people or three people in our experimental setup. The data has been collected in the same room as before, but we used two 4x16 thermal sensors appended vertically on the x-axis wall. The collected data has been processed, normalized and resized to same sizes before applying machine learning algorithms. The processed data is analyzed and used for detecting different static activities, such as standing, sitting on chair, sitting on ground, and lying on ground, by different people at the same time. Based on the experimental results, we see that Random Forest Algorithm gave us the best per-class accuracy for detecting different activities.
