Indoor Movement Prediction

Predicting user movements from temporal streams of RSS (Radio Signal Strength) measured between the nodes of a WSN (Wireless Sensor Network WSN). A static version of the jupyter notebook can be seen here.

Dataset Description

Summary

This dataset represents a real-life benchmark in the area of Ambient Assisted Living applications. The binary classification task consists in predicting the pattern of user movements in real-world office environments from time-series generated by a Wireless Sensor Network (WSN).

Input data contains temporal streams of radio signal strength (RSS) measured between the nodes of a WSN, comprising 5 sensors: 4 anchors deployed in the environment and 1 mote worn by the user. Data has been collected during user movements at the frequency of 8 Hz (8 samples per second). In the provided dataset, the RSS signals have been rescaled to the interval [-1,1], singly on the set of traces collected from each anchor.

Target data consists in a class label indicating whether the user's trajectory will lead to a change in the spatial context (i.e. a room change) or not. In particular, the target class +1 is associated to the location changing movements, while the target class -1 is associated to the location preserving movements.

The measurement campaign involved a number of 3 different environmental settings, each of which comprises 2 rooms (containing typical office furniture) separated by a corridor. A sketch of the common setup considered is provided by the Figure in file MovementAAL.jpg.

Each file in the provided dataset contains data pertaining to one temporal sequence of input RSS data (1 user trajectory for each file). The dataset contains 314 sequences, for a total number of 13197 steps. The dataset can be found here. For more information on the dataset, visit this link.

Attribute Information

Data is provided in comma separated value (csv) format.

• Input: Input RSS streams are provided in files named MovementAAL_RSS_SEQID.csv. Here, IDSEQ is the progressive numeric sequence ID. In each file, each row corresponds to a time step measurement (in temporal order) and contains the following information: RSS_anchor1, RSS_anchor2, RSS_anchor3, RSS_anchor4.

• Target: Target data is provided in the file MovementAAL_target.csv. Here, each row in this file contains: sequence_ID, class_label.

• Dataset Grouping: Data is grouped in 3 sets. File MovementAAL_DatasetGroup.csv, provides information about such data grouping. Each row in this file contains: sequence_ID, dataset_ID.

• Path Grouping: Users' movements are divided in 6 prototypical paths. File MovementAAL_Paths.csv provides information about data grouping based on path type. Here, each row in this file contains: sequence_ID, path_ID.

Project Organization

Folder Structure

.
├── data
│   ├── indoor_movement.csv
│   ├── indoor_movement_red.csv
│   └── MovementAAL
│       ├── dataset
│       │   ├── MovementAAL_RSS_1.csv
│       │   ├── MovementAAL_RSS_2.csv
│       │   ...........................
│       │
│       ├── dataset_description.txt
│       ├── groups
│       │   ├── MovementAAL_DatasetGroup.csv
│       │   └── MovementAAL_Paths.csv
│       ├── MovementAAL.jpg
│       └── README.txt
├── LICENSE
├── notebooks
│   └──  Indoor User Movement Prediction from RSS Data Set .ipynb
├── processed_data
│   ├── indoor_movement.csv
│   ├── indoor_movement_red.csv
│   └── RSS Output.png
├── .dockerignore
├── .gitignore
├── Dockerfile
├── README.md
├── docker-compose.yml
└── requirements.txt

Workflow

Dataset
      - Preprocessing
                - Concatenating Target Column with the Inputs
                - Adding Ids to Each Row
                - Adding Groups to Each Row
                - Adding Time based on 8Hz Sampling Frequency

      - Exploratory Data Analysis
                - Time Series Visualization
                - Histogram Plots
                - Dimensionality Reduction
                         - PCA (Principle Component Analysis)
                         - UMAP (Uniform Manifold Approximation and Projection)

      - Movement Classification & Prediction
               - Baseline Model
                        - Classification
                                - 75-25% Train-Test Split
                                - Classification via Random Forest
                                - Training
                        - Evaluation
                                - Validation

               - Ensemble Method for Improving Prediction
                        - Feature Extraction
                                - Tsfresh feature extraction
                                - Important feature selection
                        - Classification
                                - 75-25% Train-Test Split
                                - Classifier Ensembling Via Soft Voting (Decision Tree, KNN,
                                Gradient Boosting,Random Forest, Adaboost)
                                - 10-fold Stratified Cross Validation
                                - Training
                         - Evaluation
                                - Validation
                                - UMAP of the Extracted Features
                                - Decision Boundary Plotting

Installation

• You'll need docker and docker-compose installed on your machine

• Clone the repo

  git clone git@github.com:rednafi/indoor-movement-prediction.git

• Run the container

  docker-compose up -d

• Go to your browser and open the following URL (Your token might vary. Inspect that using docker logs indoorapp-cont command)

  http://127.0.0.1:8888/?token=9143b78935c03b423190348826bfd9194beabfdb802563b8
  

• Stop the container

  docker-compose down