Abstract
Human Activity Detection plays a critical role in the field of education. It helps to analyze and understand the non-verbal behaviors of students such as body movement and function that can help to improve the learning outcome. This thesis proposed a 1D-CNN model-based HAD method to address these key aspects. For this study, five complex human activities are selected in the lesson, such as scrolling, studying, playing games, writing on the board, and raising hands. The data for the intended activities were recorded using an accelerometer in a smartwatch. The model was trained and analyzed in Python’s programming language environment. The reported method achieved an accuracy of 97.53% using the in-house dataset for the intended activities. 1D-CNN specifically deals with the 1D data and the data size depends upon the application. Due to this, the hardware implementation of 1D-CNN is always challenging while processing the 1D data. For hardware acceleration, two hardware designs of the 1D-CNN accelerators were proposed for FPGA implementation. Both hardware designs promote parallel computation with high performance. Proposal-1 can process smaller 1D data more efficiently and offers lesser re- source utilization and dynamic power consumption in comparison to proposal-2. In contrast, proposal-2 is well suited for larger 1D data as it requires less time to complete the convolution operation in comparison to proposal-1. Proposal-1 and Proposal-2 offer 0.622 GOPs and 2.488 GOPs at a total power consumption of 0.102 W and 0.135 W, respectively. the hardware designs were implemented in a Xilinx ZYNQ-7000 FPGA.