Slobodan Dukanović presents Robust Sound-Based Vehicle Counting in Low-to-Moderate Traffic Conditions

On 2020-01-28 14:30:00 at G205, Karlovo náměstí 13, Praha 2
The analysis of traffic monitoring data enables better use of the roadway
systems (e.g. enables drivers to be better informed about traffic congestion
and parking possibilities), prediction of future transportation needs, and the
overall improvement of transportation safety. Three key functionalities of a
traffic monitoring system are vehicle counting, vehicle classification, and
vehicle speed estimation.
We address vehicle counting (VC) in low-to-moderate traffic using one-channel
acoustic sensors. The VC is viewed as a regression problem, i.e. the distance
between a vehicle and the acoustic sensor is predicted using a set of audio
features. The feature set includes the standard audio features and a new
feature, the high-frequency power, which is more robust to environmental noise
than other audio features and therefore improves the method performance in
noisy environments. The predicted distance allows for vehicle detecting/counting
via a peak detection procedure. Our strategy is to set the detection threshold
at a point where the probability of false positives and false negatives coincide
so they statistically cancel each other out in the total vehicle number. The
method is trained and tested on a traffic-monitoring dataset collected for the
purpose of this research. The dataset comprises 352 short 20-seconds mono-sound
files with a total of 1252 vehicles passing by the acoustic sensor. With the
proposed strategy, the relative VC error in traffic conditions not seen during
the training is 0.92 %.
The discussion and feedback are very welcome.
Responsible person: Petr Pošík