Slide 1

Kai Kunze and Paul Lukowicz Embedded Systems Lab, University of Passau, Insstr 43, 94032 Passau, Germany UbiComp 2007

Slide 2

Outline 1. Introduction 2. Approach Overview 3. Recognition Method 4. Experimental Validation 5. Conclusion

Slide 3

1.Introduction Present and systematically evaluate a novel method for object localization. The method provides so called symbolic location rather then absolute coordinates. The key contribution : Present a method that requires no infrastructure, relies on simple, cheap sensors. How to derived the method ? Create a mechanical excitation of the environment and analyze the response with an accelerometer and a microphone.

Slide 4

1.Introduction Two types of information can be derived from this analysis. (1). The system can be trained to recognize specific locations. (2). It can recognize more abstract locations based on materials. Advantage: Less specific positioning, the system does not need to be trained for each single location. Aims at the localization of simple objects in environments with no, or only minimal augmentation.

Slide 5

1.Introduction-Related Work Not focus on reliable, standard method. Ultrasonic Location instrumentation: BAT, MIT cricket systems. Require extensive instrumentation of the environment with ultrasonic transceivers and free line of sight and will fail to locate objects in closed compartments. Time of flight based radio frequency (RF) methods: UBISENSE ultra wide band system. Cost and effort (UWB).

Slide 6

1.Introduction-Related Work Simple Beacon Based Systems: Localization based on simple RF beacons, Bluetooth, Zigbee and WLAN, RF based system. Knowing approximate physical location can be used to constrain the search space. Indirect Localization with Sensor Signatures: Sound and acceleration. General concept of using acceleration signatures to extract location related information.

Slide 7

1.Introduction An important feature of their method is the fact that it can be used on both specific locations (e.g. my kitchen table), and abstract location types. Provide a brief description of the recognition algorithm, including, feature computation, classification, and classifier fusion. Data set contains a total of over 1200 measurements from 35 specific locations (taken from 3 different rooms) and 12 abstract location classes.

Slide 8

1.Introduction Organization On room bases (16, 9 and 10 locations) we arrive at an accuracy of between 89% and 93 % with the correct answer being in the to 2 first picks of the classifier between 97 % and 99 % of the time. With all 35 locations from the 3 rooms in one data set the recognition goes down to 81 %. However we still get the correct answer in the top 2 picks of the classifier.

Slide 9

2.Approach Overview-The Method Procedure Description (Proposed method consists of two parts): Part 1. Based on vibrating the device using a vibration-motor of the type commonly found in mobile phones. Motion and sound signals are used for an initial location classification using standard feature extraction and pattern recognition methods. Final classification is obtained through appropriate fusion of the two classification results.

Slide 10

2.Approach Overview-The Method Part 2. Based on sound sampling. Emits a series of beeps, each in a different, narrow frequency spectrum. Receives only little energy directly from the speaker. Instead a significant part of the energy comes from reflections from the immediate environment. Two parts are used together, the corresponding results are fused using an appropriate classifier fusion method.

Slide 11