Multimodal sensing-based camera applications Miguel Bordallo1, Jari Hannuksela1, Olli Silvén1 and Markku Vehviläinen2 1 University of Oulu, Finland 2 Nokia Research Center, Tampere, Finland Jari Hannuksela, Olli Silvén Machine Vision Group, Infotech Oulu Department of Electrical and Information Engineeering University of Oulu, Finland MACHINE VISION GROUP Outline Introduction • Modern movile device with multiple sensors Vision-based User Interfaces Sensor data fusion system Application case implementations • Motion-based image browser • Motion sensor assisted panorama imaging Conclusions/Summary MACHINE VISION GROUP Introduction • More and more applications and features are being crammed into handhelds • Causes usability complications given the constraints of current mobile UIs • Increased computing power not harnessed for UIs • Keypad and pointer based UIs and/or touchscreens in current devices – User’s hand obstructs the view – Require two handed operation MACHINE VISION GROUP Modern mobile device with multiple sensors • The phone includes touch screen, GPS, accelerometers, light sensor, proximity sensor • Two cameras: low resolution for video calls and high resolution for photography and video capture • Newer phones will include magnetometers, gyroscopes MACHINE VISION GROUP Motivation for vision based user interfaces Allow recognition of the context - Detect user’s actions - Recognize environment Allow 3D information Provide interactivity - Real-time feedback - Single hand use MACHINE VISION GROUP Limitations of vision based UIs Fast movements + Low light Difficult conditions MACHINE VISION GROUP The solution: sensor data fusion Fusing the data obtained by several sensors • Ambience light sensor determines illumination conditions • Video analysis detects ego-movements and analyzes the context • Accelerometers provide complementary motion data MACHINE VISION GROUP Video analysis - Every frame divided into regions - Selection of feature blocks - Estimation of block displacements - Analysis of uncertainty - Results: 4-paramenter model - X, Y, Z, r MACHINE VISION GROUP Sensor data fusion: Accelerometers MACHINE VISION GROUP Sensor data fusion Model the device movement with the folowing Define a state vector: position, speed, acceleration Define a measurement model Apply Kalman filtering adding accelerometer values: State prediction + state correction MACHINE VISION GROUP Application cases • Sensor data fusion method applied in two applications – Implemented on a Nokia N900 mobile phone • Motion based image browser – Allows browsing large images and maps with one hand operation – Works under different light conditions • Sensor assisted panorama imaging – Stitches panorama images in real time from video frames – Increased robustness against fast movements and no-texture frames MACHINE VISION GROUP Motion based image browser Uses fusion model from accelerometers + video analysis to generate commands • Scroll up/down/left/right • Zoom in/out Light sensor decides: • if camera should be turned on • weighting factors and uncertainties • 3 modes defined: • Good image quality (video analysis + accelerometer correction) • Bad image (accelerometers have increased contribution) • No image (only accelerometers are used) MACHINE VISION GROUP Motion based image browser II MACHINE VISION GROUP Sensor assisted panorama Imaging •Based on video analysis •Guides the user with instructions •>360 degrees panoramas •Real-time registration •Real-time frame evaluation and selection •Real-time frame correction •Increased robustness via sensor-data integration MACHINE VISION GROUP Panorama imaging: Sensor uses Registration •Uses sensor fusion model to compute camera motion •Increased robustness against fast movements and frames with low/smooth texture MACHINE VISION GROUP Panorama: Sensor uses II Selection •Uses accelerometer data to detect blur •Detects unwanted shake/tilt •Integrated in scoring system MACHINE VISION GROUP Summary • Vision based interfaces offer high interactivity with one hand operation • They present several limitations • Sensor fusion improves motion estimation adding robusness against fast movements and dark conditions • The framework can be included in several applications (e.g. as a part of Motion Estimation API) MACHINE VISION GROUP Conclusions • We have presented a sensor fusion framework that fuses vide analysis with motion sensors (acelerometers+magnetometers+gyroscopes) • We have presented two applications cases that make use of sensor data fusion and integration • The applications presented are by no means the only ways to apply vision or multiple sensors, and one may find new interesting possibilities in further research MACHINE VISION GROUP Thank you! Any question??? MACHINE VISION GROUP