| CPC G01N 29/12 (2013.01) [G01D 21/02 (2013.01); G01N 29/4481 (2013.01); G01N 2291/0232 (2013.01)] | 6 Claims |
|
1. A method for deriving a dangerous area of a road, comprising:
receiving, by a noise processing unit, an audio signal collected by a sensor device installed around the road;
generating, by the noise processing unit, an attenuated audio signal by removing a noise other than a noise-of-interest from the received audio signal, the noise-of-interest including at least one of a vehicle horn noise and a vehicle sudden brake noise;
detecting, by an information processing unit, the noise-of-interest by analyzing the attenuated audio signal through a learned detection model; and
establishing, by the information processing unit, a road area within a predetermined radius from the sensor device as the dangerous area of the road when an accumulated number of times the noise-of-interest is detected within a predetermined period is greater than or equal to a predetermined reference value,
the method further comprising:
before receiving the audio signal,
preparing, by a learning unit, an attenuated audio signal for training, which is a signal obtained by attenuating a noise other than the noise-of-interest including at least one of the vehicle horn noise and the vehicle sudden brake noise;
inputting, by the learning unit, the attenuated audio signal for training to a detection model that has not completed learning;
generating, by the detection model, an imitated attenuated audio signal for training that imitates the attenuated audio signal for training, by compressing and restoring the attenuated audio signal for training;
calculating, by the learning unit, a restoration loss that is a difference between the attenuated audio signal for training and the imitated attenuated audio signal for training; and
performing, by the learning unit, optimization of updating a weight of the detection model to minimize the restoration loss.
|