| Description: |
This dataset was used in [1] and donated by the authors of that
paper. Accelerometer data was collected on a smartphone installed
inside a vehicle using a flexible suction holder near the
dashboard. An expert was responsible for driving the vehicle while
the device ran an Android application called Asfault [2], developed
specifically to store the current asphalt condition continuously
over time. Asfault stores the time-stamp of the collected data,
acceleration forces in along the three physical axes, latitude,
longitude, and velocity.
The acceleration forces are given by the accelerometer sensor of
the device and are the data used for the classification task.
Latitude, longitude, and velocity are given by the GPS and are used
to locate the collected/classified data. Each one of this
information is a different continuous time series. A sampling rate
of 100 Hz was used for each time series. It is important to note
that the Android system does not guarantee a perfect precision on
the interval between readings. This means that when we choose a
sampling rate of 100 Hz, the device should output between 95 and
105 observations per second.
For this data, the (X,Y,Z) acceleration time series are converted
into a univariate time series that represents the acceleration
magnitude.
In order to obtain labeled data, the Asfault application allows the
expert to inform the pavement condition before collecting the data.
To guarantee the integrity of data, Asfault also records videos of
the road over the data collection using the built-in camera. Thus,
it is possible to perform the analysis of these videos to confirm
the class labels assigned by the expert.
The datasets were collected in the Brazilian cities of Sao Carlos,
Ribeirao Preto, Araraquara, and Maringa using a medium sized
hatchback car (Hyundai i30) and two different devices (Samsung
Galaxy A5 and Samsung S7).
The problem AsphaltObstacles involves the identification of four
common obstacles in the region of data collection. It has the
following class labels: raised_crosswalk (160 cases);
raised_markers (187 cases); speed_bump (212 cases); and
vertical_patch (222 cases). Data is variable length, minimum 111
observations, maximum 736. The data is split randomly into 50/50
default train series. The data can be resampled without bias.
The best result for the Asphalt-Obstacles dataset was achieved by
DTW distance with 81.13\% accuracy (Table 9 of [1])
[1] Souza V.M.A. Asphalt pavement classification using smartphone
accelerometer and Complexity Invariant Distance. Engineering
Applications of Artificial Intelligence Volume 74, pp. 198-211.
(Link Here)
[2] Souza V.M.A., Cherman E.A., Rossi R.G., Souza R.A. Towards
automatic evaluation of asphalt irregularity using smartphones
sensors International Symposium on Intelligent Data Analysis
(2017), pp. 322-333
|
