IEEE Technology and Society Magazine - September 2020 - 67

Cruising Detection - The Parkus Approach

Field Trial

Cruising detection is a method that discerns between
travel to a destination, and movement dedicated to
actively searching for parking. Manually identifying and
labeling the starting time of cruising for each journey
would be a time-consuming, error-prone task and is
also not practically scalable. To mitigate this problem,
the method of determining these timestamps was automated by taking into account the distance to the destination as it decreases until it hits a local minimum. At
that point the distance increases again as the driver
starts moving further away from their destination in the
search for parking. This minimum might be caused by
the driver passing by their destination and observing
that no parking is available, or it might be caused by a
circular route or S-shaped curve as might be observed
in a parking lot. This is based on the assumption that a
driver ideally would like to park as close to their destination as possible.
This constitutes the cruising ground truth, which is
used to train machine learning classifiers to automatically
detect cruising using a combination of sensor data (accelerometer, magnetometer, etc.) collected from the users'
smartphones. Overlapping time windows are used (segmentation) followed by extracting various features as
shown in Figure 1. The importance of using machine
learning to detect cruising stems from the removal of the
requirement to input a destination in order for cruising to
be detected. This allows the smartphone application to
detect cruising dynamically as users go about their everyday driving and also enables updating of the heatmap
without the need for users to provide a destination. In
addition to this, it also allows the heatmap to benefit not
only from more data, but also from more timely data as
cruising can be detected as soon as it happens rather
than at the end of a journey (since the distance to a destination curve can only be effectively generated at the journey end).
For evaluation purposes, three classifiers were trained
using the data collected from the trial including, 1) decision trees, 2) k-nearest neighbors, and 3) support vector
machines. The output of the classification step was then
post-processed using the approach of [12], in order to
reduce the effect of the classifier producing short "gaps"
of non-cruising during true cruising, or vice-versa. The
post-processing step is based on a mean filter applied to
the series of 0s and 1s representing the classifier output.
To alleviate potential bias introduced by cross validation
schemes such as in a standard k-fold cross validation
scheme, a leave-one-user-out cross validation technique
was applied for evaluation of the classifiers [13]. We also
used the dataset collected in [12] to perform several
experiments testing generalization capabilities of the
automated cruising detection system.

The main objective of the field trial was to answer the
research questions posed earlier, but in a real-life scenario. The objective was to understand the performance of
the cruise detection algorithm in a real world setting, and
reflect on its generalization capability, to better understand the impact of parking availability information on
cruising behavior and finally, to gain feedback on the
usability aspects of ParkUs.

SEPTEMBER 2020

∕

Practical Challenges and Evaluation Parameters
Signing people up to use a parking app requires that the
app provides accurate parking information. At the same
time, providing accurate parking information requires a critical mass of users since this is a crowd-sourcing approach.
Due to these conflicting objectives it was decided to restrict
the trial to a few car parks in the university where spaces
would be shared between the trial participants. Most of the
users that were recruited had fairly regular working hours,
which ensured that users arrived and departed within a certain predictable time window. This was particularly important given that this was a semi-controlled trial. Such a
setting helps to shed light on the promise of the solution as
heatmap updates from some users (those who arrive early)
benefit those who arrive later. In practice, when such a
solution goes live with a large number of participants, one
may encounter a variety of scenarios:
No heatmaps in certain places where there is no user
activity,
■■ Up to date heatmaps in places with lots of user activity, and
■■ Differences between heatmap updates and user arrivals at their destination, which could lead to variation
in the usability of the heatmap - timely and useful
when a user arrives at the destination after updates
have been made by other users in this area, and less
useful in other cases due to stale information resulting from no recent updates in the area of interest.
■■

In the worst case scenario, where heatmaps do not provide any useful information due to lack of updates, the outcome would be similar to that of the status quo (not using
any parking availability information whatsoever) and in the
best case, where up-to-date parking availability information
does exist, this will improve park search times and therefore, reduce unnecessary CO2 emissions, wastage of fuel,
time and money. In summary, there is nothing to lose from
participating in using the app. In terms of parameters to be
evaluated, we were specifically interested in the following:
■■

F1 scores, a measure of precision and recall of the
cruise detection algorithm, that highlight its generalization capability,

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

67
IEEE Technology and Society Magazine - September 2020

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - September 2020
