Jiankun Zheng (Ken)

Human Activity Recognition using Wearable Sensors by Deep Convolutional Neural Networks

• SVM method has a relatively high accuracy but a relatively high computational cost.
• Feature Selection method has a relatively low computational cost but a significantly low accuracy.
• A trade-off exists between accuracy and computational cost.
• DCNN+ and DCNN methods achieve a good balance between the two (accuracy and efficiency).

Aug/02/22

• Learned about the basics of Human Activity Recognition and gathered that 3D motion sensors often utilized accelerometers and gyroscopes send motion signals to a laptop with code ready to interpret the signals to classify the motion.
• Research on Xsens and the different modules of sensors it offered was done.

Aug/04/22

• Looked into the modules MTw Awinda (https://www.xsens.com/products/mtw-awinda) and Xsens DOT (https://www.xsens.com/xsens-dot).
• Created a document detailing the differences between the two to aid my professor in deciding which one to purchase.

Aug/06/22

• Attended weekly online meeting with UC Berkeley students and professors and exchanged our findings regarding the most suitable motion sensors for our project.
• Also discussed common actions by construction workers (i.e. excavating, laying bricks, climbing, soldering, etc.), which our final algorithm would have to be able to distinguish.

Aug/08/22

• Studied Human Activity Recognition using Wearable Sensors by Deep Convolutional Neural Networks and discussed this paper with my professor.
• Did research and watched lecture videos on common libraries implemented by machine learning programs, such as SciKit-Learn (https://www.youtube.com/watch?v=0Lt9w-BxKFQ&t=904s), NumPy (https://www.youtube.com/watch?v=zoLHuhefMNw), MatPlotLib (https://www.youtube.com/watch?v=3Xc3CA655Y4), etc.

Aug/10/22

• Coded an iris flower-distinguishing program using Sklearn.
• Began to code and train a standing-or-walking-distinguishing program as a first step to our final program to be implemented at construction sites.

Aug/12/22

• Successfully read in all the data from the UCI HAR Dataset (https://archive.ics.uci.edu/ml/datasets/human+activity+recognition+using+smartphones) into my machine learning program.
• Studied an open source HAR algorithm on GitHub (https://github.com/ma-shamshiri/Human-Activity-Recognition/tree/main/code) to familiarize myself with typical HAR programs.

Aug/14/22

• Successfully trained my program using the UCI HAR Dataset, but prediction accuracy was not ideal.
• Experimented with common HAR classifiers seen on GitHub to try to find better prediction accuracy.