Download data here: fall-dataset
In the context of the project SHAPES, in ARCO Research Group we have developed a real-time fall detection system. In order to accomplish this task, it was needed to create a dataset to train our machine learning algorithm. A lot of dataset for fall detections has been developed along the time, but none of this fit in our case of use, because the sensors were not those available to us or the sensors were not placed on the location we needed. For this reason, it was decided to develop our own dataset with the following details:
- The sensor is located on the waist.
- The raw data collected are accelerations in g, rotations in º/s and absolute orientation in Euler angles.
- A mat has been used to collect the data.
17 subjects have performed 9 exercises divided between Falls and ADLs to build this dataset. On fall-dataset you can find CSV files with all the data collected, one file for each subject, and each row is labeled to identify if is a fall or an ADL (0=ADL, 1=Fall). The CSV files are divided in three sections:
- fall-dataset-features: Each row of this dataset contains the features used in our study to filter raw data and describe a movement. Each row represents a complete exericse (Fall or ADL)
- fall-dataset-raw: Raw data from a one second window when the user perfomed the activity. Each row alone is not relevant because it only contains raw data in a instant of time. In order to get relevant information you must use all the data with the same value on the column index, all this data are part of the same exercise along the time.
- fall-dataset-all: On these files, all the data collected when the exercises were performed by the users is saved. It could be useful if you need data out of the one second window. This data is not labeled, but you can use fall-dataset-raw in order to find when a fall or an ADLs were produced. Both fall-dataset-raw and fall-dataset-all have timestamp in order to ease this task.
To conclude, it is important to say that because of some limitations on the sensor used, the amount of raw data collected on each exercise can vary.
Here it is presented some of the most important characteristics of the subjects used to collect the data.