Audio Location Accurate Low-Cost Location Sensing.pptVIP

Audio Location Accurate Low-Cost Location Sensing Overview of talk Audio for fine-grained location Prototype: detecting human sounds Evaluation Application area: 3D user interfaces Background Fine-grained location systems have been built using ultrasound e.g. Bats (ATT), Cricket (MIT) Achieving ~3cm 3D accuracy 95% of the time Many end-user devices have audible-range I/O built in, but few have ultrasound Can we use integrated/off-the-shelf audio hardware for location? Audio-based location withoff-the-shelf hardware Many tx Speakers in environment as tx Mobile phones or PDAs as rx Privacy-preserving Many rx Computer microphones as rx Can use mobile devices for tx BUT can also use human sounds for tx Locating human sounds Need coverage from at least 4 mics Unknowns: X,Y,Z,t (t=time of sound) More is better since occlusion happens Users do not need special “tag” No per-user setup required Lowers costs and increases simplicity User identity is not provided Many apps do not need identity Anonymity is good for privacy Could fuse with identity e.g. from RFID Aims of prototype Fine-grained location sensing with hardware accessible to end users What accuracy can we obtain for locating human sounds, e.g. finger clicking, hand clapping? Application area: 3D user interfaces using human sounds Prototype Use standard PC Add 6 PCI sound cards and 6 mics Total cost of sound hardware ~￡100 From Fedora Core 2 Linux distribution Java software System Architecture Signal Detection Problem: identify the same part of the same sound in audio streams from multiple mics Amplitude-threshold algorithm Keep track of current noise floor Mark sample as “significant” when amplitude is is at least F times noise floor. (F ≈ 2.5) Properties Very good at detecting sharp sounds Equally important: ignores other sounds Robust to noisy environments and cheap mics Timing Need time sync for sound streams 1ms error ≈ 30cm in space Problem: Linux/Java introduce delays Buffering and scheduli