Signal Processing Feature Extraction in regards to Siren Detection

Hello Everyone,

I am writing a code to simulate siren detection. What I want to do is to extract features from the sirens so that my code can distinguish them from horns, construction, and other roadway noise.

I need help with ideas for features to extract. We've tried Maximum Power, Total Energy, Max power index change (to show oscillation), and a few others, but unfortunately, our histogram plots overlap too much for these to be viable features.

Any help or tips in the project would be greatly appreciated, we plan on having the matlab simulation working by the end of February, and then work on implementing to DSP. Thanks!

[Hamza Ashraf]

hi sir am working on the same problem can you help me out with the code

i would be gratefull if you could do this

[siva MADDURI]

First we process the siren sound and ambulance light detection in survillence camera by video processing...By combining both results we confirm the ambulance present in traffic at a distance.

Answer 1

[Image Analyst]

Do you have a set of known siren "types" that you can use to train the system? Maybe you can use deep learning to recognize sirens if you gave enough sample waveforms for it to figure out what audio characteristics are characteristic of sirens in general.

Answer 2

[Stephan]

Hi,

your question is nearly unanswerable. You ask very unspecific and we do not know what you already did. What kind of sirens do you want to detect? German police and american emergency ambulance sound different. So what is the aim here? If you for example look to the germans, there are different sounds for all the different organisations. Then those sounds are different in many cases. For example the classic "Martinhorn" with pressure air or an electronic version? There are also different versions (frequencies) for city or country, due to the audibility of the different frequencies in different environment.

Normally, the sirens of the fire department in germany are tuned to beating. Is it the similar where you are looking for an approach? This could be a starting point for you! What about the fft? Why not search for a specific pattern in the frequencies of the audio signal? Look at the link i gave you: Martinhorn for german fire trucks has specified frequencies you can look for...

I give you an example of an typical german fire truck, which is attached as a mat-file and as audio file. Also you find the code which leads to the diagram of the correspondig fft-analysis below:

L = size(data,1);
Y = fft(data);
P2 = abs(Y/L);
P1 = P2(1:L/2+1);
P1(2:end-1) = 2*P1(2:end-1);
f = fs*(0:(L/2))/L;
plot(f(1:20001),P1(1:20001))

You can see the beat and the harmonic frequencies. I think this is useful for your purpose if you adapt it to the case you have.

Best regards

Stephan

[Image Analyst]

Have you ever seen the deep learning demo where they found images of hammers? You'd be totally surprised at how accurate it is in finding a huge, huge variety of VERY different looking hammers.

[Stephan]

I did not see this. But surely you are right. If there are enough samples for learning, impressive results can be reached.

What i wanted to point out was, that i would try to use (also) the frequency based data for this purpose, since the sirens will look very different to other sounds. They have some kind of fingerprint in this representation of data. This will help to classify.