Terrain Identification using Sensory Prosthetic Limb Timeseries Data
ECE 542 (Neural Networks) Project 1
Group Project (3 members)* — (See My Contribution) — Status: Complete
Sept 2020 - Oct 2020 — Raleigh, North Carolina
Overview:
Prosthetic limbs are required to perform a variety of functions depending on how they are being used. For example, prosthetic legs can be used for walking, running, climbing, traversing difficult terrain, stairs, and more. In order to adjust to different terrain to perform these disparate functions, the terrain must be identified by some means. In this project, accelerometer data from an individual walking on a flat surface, walking on grass, and going up/down stairs is used as a timeseries training data for a deep learning model to identify the terrain being traversed. This was a project for ECE 542.
Using Keras and the Google Colab environment, I ran 16 different neural network architectures in parallel to find the best performing model in identifying the terrain an impaired individual with a prosthetic limb was overcoming. Afterward, a simple gridsearch was performed for hyperparameter tuning. I used MATLAB for all the data pre-processing and post-processing. As seen in the report, a multi-layered convolutional network for the timeseries data was preferred as it was able to extract features from the data not normally seen. Afterward, a recurring network was used in a small window to evaluate said features and label the terrain accordingly. The data was overwhelmingly skewed torward the 'walking on flat surface' terrain, which is why in the report it's seen some predictions oscillate between other terrains and the previously stated. a SMOTE could be incorporated to improve this discrepency, however, its proven more difficult to incorporate such concepts due to my (current) lack of experience with SMOTE + timeseries data.
Despite this, one of my models obtained an F1 score of 83.8%, which is 3rd in the class (of 28 teams).
My teammates only decided to help me last second before the deadline, so they only helped with writing the report...
Similar to Semantic Segmentation of Crop Damage, I unfortunately did a disproportionate amount of the work. I completed the entire project by myself, but I fortunately had a little bit of help on the report.
