Habitat Telemonitoring System Based on the Sound Surveillance

This paper presents a telemonitoring system in an habitat equipped with physiological sensors, position encoders of the person, and microphones. The originality of our approach consists in replacing the video camera monitoring, not well accepted by the patients, with microphones acquiring the sounds. The sounds are analyzed and not stored in order to maintain the person privacy. We present the entire telemonitoring system which makes the data fusion between medical information and sound information and particulary the sound processing algorithms to detect a distress situation. The first step of sound processing is the sound event detection in a noisy everyday life environment. Sound event detection is necessary to extract the significant sounds before initiating the classification step. Sound classification system and its performances are presented in this paper, too. Introduction Medical monitoring is more and more frequently used in order to reduce hospitalisation costs. There are many researches in telemedicine, but few of them are sound based. In this paper, we present a medical telemonitoring system with a smart audio sensor. The system we work on is designed for the surveillance of the elderly, convalescent persons or pregnant women [1]. Its main goal is to detect serious accidents as falls or faintness at any place in the apartment. It was noted that the elderly had difficulties in accepting the video camera monitoring, considering it a violation of their privacy. Thus, the originality of our approach consists in replacing the video camera by a system of multichannel sound acquisition. The system analyzes in real time the sound environment of the apartment and detects the abnormal sounds (falls of objects or patient) and the calls for help, that could indicate a distress situation in the habitat. Again, to respect privacy, no continuous recording or storage of the sound is made, since only the last 5s of the audio signal are kept in a buffer and sent to the alarm monitor if a sound event is detected. The sound information extraction is a complex task because the audio signals occur in a noisy environment and the everyday life sounds are extremely diverse. We shall start with an overview of the medical telemonitoring system, next we shall expose the detection algorithm necessary to extract the sound event and we shall finish with the classification system. The Sound Analysis System The habitat we used for experiments is a 30 m apartment situated in the TIMC laboratory buildings (See the Figure 1 (a)), filled with various sensors, especially microphones. A microphone is placed in every room (toilet, kitchen, shower-room, hall and living-room). This allows a continuous sound surveillance in the entire apartment. Each of the 5 microphones is connected to the slave computer. The sound or speech source can be localized. Figure 1 Smart Habitat plan and analysis system The microphones used are omni-directional, condenser type, small size and low cost. A signal conditioning card, consisting in an amplifier and an anti-aliasing filter is associated to each microphone. The acquisition system consists in a multi-channels acquisition card PCI 6034E of National Instruments, used with a 16KHz sampling rate (usual frequency in speech applications). The sound analysis has two steps: the first step concerns the detection of a sound event and the second one the sound classification (see Figure 1 (b)).The second step includes an automatic sound classification and a recognition of calls for help expressions. In the first step, signals from all the 5 channels are used to detect events. It is a difficult task because of the environmental noise. If a sound event is detected, extracted signal is transmitted to the second step and sound classification is initiated. The everyday sounds are divided in 9 classes. The criteria used for this repartition were : ?? Statistical probability of occurrence in everyday life ?? Possible alarm sounds (scream, person fall) are priority ?? The duration of the sound: significant sounds are considered to be short and impulsive The 9 sound classes are divided in 2 categories: normal sound classes (door clapping, phone ringing, step sound, human sounds (cough, sneeze,...), dishes sound, door lock) and sound classes that generate an alarm (breaking glasses, screams, fall sounds). In conclusion, if an abnormal sound class is detected or a call for help is recognized the sound analysis system transmits an alarm to the data fusion system (which fusions sound and medical sensor information). The decision to call the emergency is taken by the data fusion system. The entire telemonitoring system is composed of two computers which exchange information through a CAN bus (see Figure 2). The CAN bus is low cost and provides a high level of security (only sensors are connected on the bus). Figure 2 The acquisition and analysis system Slave PC Signal conditioni ng