Title: Activity Recognition with Radio-Frequency Identification in Cognitive Assistive Robotic Environments

Author: Ronnie Smith (@ronsm)

Submitted: 16th August 2019

This repository stores all of the code developed during the project. The software within the Training directory is a collection of standalone module that can be run (in a specific order, see below) to generate an LSTM or ConvLSTM model from raw RFID data gathered from an IMPINJ Speedway Revolution system. Within the Deployment directory you will find a single application, comprised of a number of modules (including re-worked versions of those from the training stage), that combines the previously generated LSTM or ConvLSTM model and some ontological knowledge to make coarse-to-fine grained predictions of activities.

To understand fully the purpose and design of this software, you may first wish to read the accompanying thesis, available on my website: https://ronsm.com/courses/mscr-project/thesis.pdf

The following modules are provided within the Training directory:

• Data Collector Module
• Data Segmentation Module
• Data Converter Module
• Classification Module

There are individual readmes for each module. Here we will simply discuss some parameters and the order of execution of the modules to ultimately generate an LSTM or ConvLSTM model.

This module must be run first, to collect the data that will be used by all of the other standalone modules.

See the README inside this module's directory for information on how to setup and use it.

Using the label that you have previously provided to the DCM, the Data Segmentation Module can help you automatically split the original collection, which contains many activities, into a single collection for each activity.

You must run this module before attempting to run either the Data Converter Module or the Classification Module.

See the README inside this module's directory for information on how to setup and use it.

The standalone Data Converter Module needs to be run before the Classification Module, since it requires a dataset to be provided to it from DCvM.

It will generate a dataset inside its own directory (titled dataset) which you should copy in its entirety to wtihin the root directory of the Classification Module - overwriting a previous 'dataset' folder if need be.

See the README inside this module's directory for information on how to setup and use it.

Once the files have been provided from the DCvM, the Classification Module will generate and evaluate numerous LSTM or ConvLSTM models (ConvLSTM is the default).

The best performing model will be saved within the root directory of the module as best_model.h5. This file is required by the sofware in the Deployment directory to make predictions. Place the best_model.h5 file into the Deployment/models directory - again, overwriting a previous model if need be.

See the README inside this module's directory for information on how to setup and use it.

The software inside the Deployment directory is intended for use once the training phase has ended. It contains two pieces of software: a Data Collector Module that is designed to run continuously to gather fixed-size windows of data, and a control module that orchestrates the entire activity prediction process.

See the README inside this module's directory for information on how to setup and use it.