We are developing an application that will alert unaware pedestrians and those with hearing disabilities if a car is approaching them. Has applications ranging from helping reduce the no. of accidents occuring due to pedestrians carelessly crossing the streets with earphones on, alerting players of VR/AR games (for e.g., Pokemon Go) of oncoming cars to allowing people with aural impairments cross the streets with ease. We call it WalkSafe.
WalkSafe lets you walk freely on streets and everywhere else in general, even when you are wearing earphones or have difficulties in hearing.
Refer to
Dataset Preparation.ipynbto jump to the code directly.
We are using an open dataset (titled Urban8K) containing various urban sounds including car horns. We also use about an hour long recording of urban background noise to generate unique positive examples for training our model. We cropped out individual horns from the dataset and randomly superposed them on 2 second clippings from the background noise to create 1800 positive and 700 negative training examples. The figure given below shows how the individual bursts of horn were cropped. The cropped out horns were about 0.4 seconds long and were randomly superposed over 2 seconds long clips of background noise.
| Original | After cropping |
|---|---|
 |
 |
The dataset was then labeled by creating an array having 10 1s starting right
from the point where the horn clip just ends and zeros elsewhere.
This method usually gives good results since we term a sound to be horn only after
it has ended. This perfectly aligns with the fact that one cannot predict a sound to be horn by only listening to a really miniscule sample of it (1 ms in this case!)
The labels were generated along with the training examples and were saved in a Numpy pickle.
Refer to
graph_spectrogramfunction inutils.pyto jump to the code directly.
Audio data, if used raw, doesn't come with a good set of usable features. One has to
pre-process the raw audio signal to make it suitable for training purpose. All the modern microphones use 44.1 KHz sampling rate to record the audio. If we load the raw audio (which is 2 seconds long) directly, we therefore get a vector of shape . This is a whooping lot of features!
We can reduce the no. of features drastically by switching to a more suitable domain, i.e., frequency domain. There are a large no. of frequencies in a typical audio stream. The figure given below shows how one is likely to have a lot of features in the time domain whereas, in frequency domain, the number of them gets significantly reduced due to the discrete nature of frequencies. A signal in time domain can be converted to frequency domain by calculating Short-time Fourier Transform over the audio signal.
 |
|---|
| Time Domain Vs Frequency Domain |
We chose a window size of 1650 time steps, step size of 65 time steps and fft length to be 200 after multiple trials.
(More details about these parameters are given in the docstring of graph_spectrogram function inside the utils script)
This provides us back with a matrix of shape .
This is a lot of reduction in the parameters, plus frequency domain is the most suitable domain
for dealing with signals and always gives better results than the time domain. Thus,
finally we end up with a highly useful, yet computationally efficient set of parameters.
 |
 |
 |
|---|---|---|
| Aniket Teotia | Suyog Jadhav | Udbhav Bamba |