Infant Crying Detection Under Monitoring Scenario

Baby crying is the main way of communication with the outside world,and it is also a prototype of human early language.Early babies are in sleep for most of the time,and if cries are monitored during the baby's sleep and a real-time reminder is given,it can reduce a lot of unnecessary companion time,which is of great practical significance to alleviate the guardian's nursing pressure and psychological burden.At present,there are some unattended systems in the sleep state of infants used in combination with cameras or sensors,which can provide an early warning when the baby wakes up and cries.However,users need lots of money and energy to purchase and use additional equipment.Compared with the high recognition rate of video surveillance and motion capture,the system with only audio monitoring generally has a low recognition rate.In order to reduce the user's usage threshold and cost,this paper proposes a monitoring system based on mobile device application,combining mainstream deep learning and pattern recognition methods,and providing dynamic update and extended terminal recognition capabilities.It is more convenient for families to use these smart care systems with low-cost and relatively efficient babies.The work mainly includes the following parts:1.Create a database of baby crying in the monitoring scene.The data comes from freesound.org.The main data is the audio data recorded by individual users in the indoor environment using a mobile phone or other devices.Four major data categories were collected through user-defined labels and classification information: baby crying,baby laughter,background white noise,and ambient noise.2.GMM Fisher Vector based baby crying audio modeling method to solve the feature vector alignment problem caused by the different length of crying sound.In the past,processing feature alignment was usually to intercept the same length of audio or directly use the feature mean,resulting in the loss of detail distribution of timing features and change information.GMM Fisher Vector can extract the local features from equal audio and maximize the retention of feature details to improve the stability and reliability of the model.3.Deep learning based baby crying recognition model,which can better learn and detect some potential features in audio data compared to traditional audio recognition models(VQ,DTW,GMM,HMM,RFC,SVC).It is more inclusive and understandable for data expansion and feature complexity.4.Established a mobile baby crying recognition system.It uses the Mel Cepstrum Coefficient(MFCC),which can sense the hearing characteristics of the human ear,combined with short-term zero-crossing rate(ZCR)and short-term energy mean square(RMSE).Class feature,Combined with GMM Fisher Vector and DNN model,it can effectively detect baby crying and alarm,and has offline recognition and online update function.