This project is aimed to detect the baby locked in the car by careless parents. Several methods can be used for detection, for instance, sensing the pressure of the baby seat; monitoring the condition inside the car or recording voice inside the car near the baby’s seat. The first method is easy to implement, since we only need to install a tabletting resistance at the bottom or backrest of baby seat and if there’s force on the sensor, we can tell that there’s a baby on the seat/car. However, this scheme does not work if the baby or small child move out of seats. Another approach is monitoring the sight inside the car, which is more reliable comparing to pressure sensor. However, this scheme needs more computation since CV(computer version) algorithm is more complex. Another important disadvantage of this approach is the camera cannot work well if the light inside the car is too dim.
The approach we used is recording the sound inside the car and detecting whether there’s baby crying sound. If a baby/small child is locked in the car, he/she will cry loudly, which can be recorded by microphone. The surrounding inside the car is relatively closed, thus there’s not so much noisy sound that detecting baby’s crying is relatively simple.
Raspberry Pi inside the car would record the surrounding sound every few seconds, and every piece of recorded sound is 3 seconds length. After that, python feature extraction API is used to extract 160 specific sound features from the latest record sound. These features include statistics information of MFCC(Mel-frequency cepstral coefficients), energy and SSC(Spectral Subband Centroid).
The approach we used is recording the sound inside the car and detecting whether there’s baby crying sound. If a baby/small child is locked in the car, he/she will cry loudly, which can be recorded by microphone. The surrounding inside the car is relatively closed, thus there’s not so much noisy sound that detecting baby’s crying is relatively simple.
Raspberry Pi inside the car would record the surrounding sound every few seconds, and every piece of recorded sound is 3 seconds length. After that, python feature extraction API is used to extract 160 specific sound features from the latest record sound. These features include statistics information of MFCC(Mel-frequency cepstral coefficients), energy and SSC(Spectral Subband Centroid).